NCP-GENL Practice Test 16 — 15 Questions

Question 1

A team has built a TensorRT-LLM engine for a 70B model on four NVIDIA H100 GPUs using tensor parallelism. They now need to serve the same model on a single H100 for a development environment, accepting higher latency. What is the most appropriate approach?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Rebuild the engine with tensor parallelism set to one and apply weight quantization so the model fits in a single GPU's memory.. Tensor parallelism is a build-time property, so changing the number of GPUs requires rebuilding the engine with tensor parallel size one. Because a 70B model in FP16 exceeds a single H100's memory, weight quantization is also needed to fit. The other options either attempt to reuse an incompatible engine, misapply parallelism concepts, or increase precision, none of which yield a working single-GPU deployment.

Answer

Use pipeline parallelism with a single stage to split the model across the GPU's SMs automatically.

Answer

Load the existing four-GPU engine on a single H100 and set the runtime to use only one device.

Answer

Convert the engine to a TensorRT plan file with FP32 precision so it can use unified memory on the H100.

Question 2

An ML engineer is fine-tuning a 7B-parameter model with LoRA on a single NVIDIA A100 40GB GPU. The training script reports that the adapter weights are not updating after several hundred steps, and the loss remains flat. The base model weights are frozen as intended. Which LoRA configuration issue is the most likely cause?

Accepted Answer

The target modules list is empty or points to modules that are not part of the model's forward pass.. If the LoRA target modules list is empty or references modules that are not executed during the forward pass, the adapter parameters receive no gradients and never update. The base model remains frozen as intended, but the adapters are effectively absent from training, which explains both the flat loss and the unchanged adapter weights.

Answer

The base model is loaded in 8-bit precision, which prevents any gradient computation on the adapters.

Answer

The learning rate is too low, so the adapter weights change by amounts below floating-point precision.

Answer

The LoRA rank is set too high, causing the adapters to be initialized to zero.

Question 3

You are building the data preparation stage for an NVIDIA NeMo retrieval-augmented generation pipeline that will ingest millions of internal wiki pages. The ingestion team reports that the same policy text appears in dozens of pages with minor edits, and that some pages contain copied tables from external sources. You need to produce a clean, deduplicated chunk store that supports accurate citation and avoids returning redundant passages. Which TWO actions best address these requirements? (Choose two.)

Accepted Answer

Run near-duplicate detection at the chunk level using MinHash or SimHash, and retain one canonical chunk per duplicate group while preserving a mapping from the canonical chunk to all source page identifiers.. Redundant retrieval from repeated policy text is best solved by chunk-level near-duplicate detection that keeps one canonical copy while mapping it to all source pages, and accurate citation requires source metadata to travel with each chunk through the embedding and indexing stages. Together these actions reduce duplicate results and preserve provenance, which fixed-size chunking, page-level embedding, or blanket exclusion of external tables cannot achieve.

Answer

Apply a fixed-size token window with no overlap to every page so that chunk boundaries are uniform across the corpus.

Answer

Tag chunks that contain tables copied from external sources as low-quality and exclude them from the index entirely.

Answer

Embed every page as a single vector and store the full page text as the retrieval unit, avoiding chunking entirely.

Question 4

A team is evaluating a large language model (LLM) deployed with NVIDIA Triton Inference Server for a real-time question-answering system. They observe that the model's responses are accurate but latency spikes during peak load, causing timeouts. They need to evaluate the model's performance under high concurrency to identify the maximum throughput while maintaining a 95th percentile latency below 200 ms. Which evaluation approach should they use?

Accepted Answer

Use NVIDIA Triton's Performance Analyzer with concurrency sweep and latency constraints.. Evaluating performance under high concurrency requires simulating multiple simultaneous requests and measuring latency percentiles. NVIDIA Triton's Performance Analyzer is designed for this, allowing concurrency sweeps and latency constraints to find the maximum throughput meeting the SLA. Other options either lack concurrency or measure irrelevant metrics, so they cannot identify the operating point that satisfies the 95th percentile latency requirement.

Answer

Measure GPU utilization with NVIDIA Management Library (NVML) during a single long-running inference.

Answer

Run a one-off inference with a batch size of 1 and measure latency.

Answer

Compute the model's FLOPS and compare against the GPU's peak theoretical FLOPS.

Question 5

An engineer is tuning a TensorRT-LLM deployment of a 7B model for a latency-sensitive API. Profiling shows that time per output token is higher than expected and that many small kernels run back to back with gaps between them. Which TWO changes are most likely to reduce the per-token latency by cutting kernel launch overhead and redundant memory traffic? (Choose two.)

Accepted Answer

Enable CUDA Graphs so the decoding iteration is captured and replayed as a single graph launch.. The profile shows many short kernels with idle gaps, which is the classic signature of launch overhead and unfused elementwise work. CUDA Graphs collapse the decode iteration into a single replayable launch, and fused attention and GEMM plugins merge operations that would otherwise round-trip intermediates through HBM. Together they reduce both launch count and memory traffic, which are the two costs identified in the profile.

Answer

Enable a debug synchronization after every kernel so profiling timings are accurate.

Answer

Raise the maximum sequence length in the engine build to accommodate the longest possible prompt.

Answer

Switch the model weights to FP32 to avoid any dequantization kernels in the decode path.

Question 6

A production LLM inference service runs on NVIDIA Triton Inference Server across multiple GPUs. The SRE team wants to detect when the service starts returning incorrect or degraded responses compared to a baseline, even when latency and throughput remain normal. Which monitoring approach is most appropriate?

Accepted Answer

Implement model output quality checks by comparing responses against a baseline using statistical or embedding-based metrics.. To detect degraded or incorrect responses while latency and throughput are normal, monitoring must focus on output quality. Comparing model outputs to a baseline using statistical or embedding-based metrics can reveal semantic drift or accuracy drops. Resource and performance metrics like GPU utilization, error rates, and latency are insufficient because they do not reflect the correctness of generated text.

Answer

Monitor GPU utilization and memory usage via NVIDIA DCGM to detect anomalies.

Answer

Use NVIDIA Nsight Systems to profile the inference pipeline and identify bottlenecks.

Answer

Set up alerting on Triton Inference Server's error rate and request latency percentiles.

Question 7

A developer is creating prompts for an NVIDIA NIM-hosted LLM to summarize financial reports. The reports are lengthy and contain many tables and figures. The developer wants to ensure the summaries are accurate and include key numerical data. Which TWO prompt engineering techniques should be applied? (Choose two.)

Accepted Answer

Instruct the model to extract and include all monetary values and percentages exactly as they appear in the report, and to avoid rounding or paraphrasing numbers.. To ensure accurate financial summaries with key numerical data, the developer should explicitly instruct the model to include all monetary values and percentages exactly as they appear, and provide a few-shot example demonstrating correct inclusion of numbers. These two techniques directly guide the model to preserve numerical accuracy and follow the desired format. High temperature, chain-of-thought, or ignoring tables would not achieve the goal.

Answer

Use a chain-of-thought prompt that asks the model to first list all tables and figures, then write the summary.

Answer

Set the temperature to a high value like 1.0 to encourage the model to creatively interpret the financial data and provide insightful analysis.

Answer

Instruct the model to ignore any tables and figures and focus only on the narrative text to avoid confusion.

Question 8

A team is fine-tuning an 8B-parameter LLM with LoRA on a single NVIDIA A100 80GB GPU using NVIDIA NeMo. They observe that training loss decreases, but validation loss starts to rise after epoch 2. They want to keep the same dataset and hyperparameters but mitigate overfitting. Which change is most appropriate?

Accepted Answer

Reduce the number of training epochs and apply early stopping based on validation loss.. The divergence between decreasing training loss and increasing validation loss after epoch 2 is a textbook overfitting signal. The most direct, minimal-risk remedy is to stop training earlier using validation loss as the criterion, preserving the checkpoint that generalizes best. Enlarging adapter capacity or altering optimizer and batch settings does not target the generalization gap and may worsen it.

Answer

Increase the batch size and scale the learning rate proportionally to speed convergence.

Answer

Switch the optimizer from AdamW to SGD with momentum to regularize the adapter.

Answer

Increase the LoRA rank from 8 to 64 to give the adapter more capacity.

Question 9

A team fine-tunes a model with NVIDIA NeMo using a packed sequence dataset and notices that some training samples contain several short conversations concatenated. They must ensure the loss is computed only on assistant responses and not on the packed boundaries. Which configuration detail should they verify?

Accepted Answer

That the loss mask aligns with each sample's response tokens and that packed sequences are separated by an attention boundary. Packed sequences boost training efficiency by filling each context window with multiple samples, but correctness depends on the loss mask covering only assistant response tokens and on attention being blocked across sample boundaries. Batch size, vocabulary changes, and gradient accumulation do not affect these two requirements.

Answer

That the global batch size is increased to compensate for the additional tokens in each packed sequence

Answer

That the tokenizer vocabulary is expanded to include special separators between packed samples

Answer

That gradient accumulation steps equal the number of samples packed into each sequence

Question 10

An engineer fine-tunes a model on a domain corpus with NVIDIA NeMo and observes that training loss falls steadily while validation loss begins rising after the second epoch. The team must produce the most generalizable checkpoint without changing the dataset. Which action should they take?

Accepted Answer

Enable early stopping based on validation loss and restore the best checkpoint. When training loss keeps falling but validation loss turns upward, the model is overfitting and the best generalizing weights occur near the point where validation loss is lowest. Early stopping with best-checkpoint restoration captures that point. More epochs, a higher learning rate, or dropping validation all fail to address the divergence or actively make it worse.

Answer

Increase the number of training epochs so the model fully converges on the domain corpus

Answer

Disable validation during training and rely on the final training loss to select the checkpoint

Answer

Raise the learning rate to escape the local minimum the model has settled into

Question 11

A developer is using an NVIDIA NIM-hosted model to classify support tickets into a fixed set of categories. The model occasionally invents new category names. The team wants to guarantee that only allowed categories are returned. Which approach is most appropriate?

Accepted Answer

Provide the full category list in the prompt and use the NIM guided_choice parameter with those categories.. When output must be one of a fixed set, enumerating the categories in the prompt and applying a decoding constraint such as guided_choice ensures only valid labels are emitted. Few-shot examples and reasoning improve quality but do not enforce the boundary. High temperature and broad top_k work against the requirement by increasing variability.

Answer

Use few-shot examples of each category and set top_k to 100.

Answer

Ask the model to explain its reasoning before choosing a category.

Answer

Add 'Choose from the list' to the prompt and set temperature to 1.0.

Question 12

A developer needs to fine-tune a 7B LLM for a customer-support chatbot using NVIDIA NeMo. The dataset contains paired instructions and desired responses. Which data format should be used to prepare the dataset for supervised fine-tuning?

Accepted Answer

A JSONL file where each line contains an input text field and an output text field representing the instruction-response pair.. Supervised fine-tuning trains a model to produce a target response given an input instruction. The dataset must therefore contain aligned input-output pairs, and NeMo's data pipeline accepts JSONL records with distinct input and output fields. Files lacking response labels, pre-tokenized sequences without pairing, or pure configuration files cannot supply the supervision signal the training objective requires.

Answer

A YAML configuration file listing hyperparameters and dataset paths only.

Answer

A pickle file containing a Python list of tokenized integer sequences with no text labels.

Answer

A CSV file containing only the raw customer queries with no corresponding responses.

Question 13

An engineer is designing prompts for an NVIDIA NIM-hosted model that must extract structured fields from unstructured invoices. The extraction accuracy is inconsistent across vendors with different layouts. Which TWO prompt engineering techniques would MOST improve reliability? (Choose two.)

Accepted Answer

Instruct the model to output fields in a fixed JSON schema and to use null for any field not present in the document.. Reliable structured extraction combines format constraints with demonstrated patterns. A fixed JSON schema with a null convention makes outputs machine-parseable and defines behavior for absent fields, while few-shot examples across varied layouts teach the model to handle layout diversity. Together they reduce both structural errors and hallucinated values, which prose summarization and high-temperature sampling would worsen.

Answer

Ask the model to summarize the invoice in prose first and then extract fields from its own summary in a second pass.

Answer

Raise the temperature to 0.9 so the model explores multiple interpretations of ambiguous invoice text before committing to values.

Answer

Remove all formatting and concatenate the entire invoice text into a single lowercase string before sending it to the model.

Question 14

You are building a pretraining dataset from a large collection of source-code repositories for an NVIDIA NeMo LLM. The data includes many files with licenses, generated code, and minified JavaScript. Which NeMo Curator-based approach best improves code data quality before tokenization?

Accepted Answer

Use language-specific filters that detect and remove minified files, license headers, and auto-generated code, then apply deduplication.. Language-specific filters combined with deduplication directly target the described quality issues. Minified files, license headers, and generated code are common in code corpora and can be detected with heuristics such as average line length, ratio of whitespace, and presence of standard license text. Deduplication removes repeated generated files. This preprocessing ensures the pretraining corpus contains meaningful code, which improves the model's code capabilities.

Answer

Convert all code to a single programming language using an automated transpiler before tokenization.

Answer

Increase the model's context window to 32k tokens so that entire repositories fit in one sequence.

Answer

Tokenize all files with a byte-level BPE tokenizer and skip any further filtering.

Question 15

A healthcare AI team is evaluating a large language model for clinical note summarization. They want to measure whether the generated summaries contain fabricated information not present in the source notes. Which evaluation approach is most appropriate for detecting hallucinated content?

Accepted Answer

Use a natural language inference (NLI) model to check entailment between the source note and the generated summary.. Natural language inference (NLI) is effective for hallucination detection because it evaluates whether the generated summary is logically entailed by the source document. If the summary introduces unsupported information, the NLI model will not predict entailment. Other metrics like perplexity, BLEU, and ROUGE-L focus on fluency or surface overlap and cannot reliably identify fabricated content. Therefore, NLI-based checking is the most appropriate approach.

Answer

Compute the perplexity of the generated summaries on a held-out set of clinical notes.

Answer

Calculate the BLEU score between the generated summary and a reference summary written by a clinician.

Answer

Measure the ROUGE-L score between the generated summary and the source note.