Software Development — Python, Frameworks and Orchestration — NVIDIA NCA Gen AI LLMs practice questions

The clearest distinction between a chain and an agent is that an agent:

1. Runs on GPU while chains are restricted to CPU execution
2. Can only be used with retrieval-augmented generation workloads
3. Must be fine-tuned on tool-use data before it can call any tool
4. Decides at runtime which tool to invoke and when to stop

Answer: D — Decides at runtime which tool to invoke and when to stop

D) Correct — the defining property is model-driven control flow: the agent selects actions and termination. A) Hardware placement is orthogonal to both patterns. B) Agents are general and RAG can be built as a plain chain. C) Many capable instruct models call tools well without task-specific fine-tuning.

The PRIMARY job of a collate function in a PyTorch DataLoader for LLM fine-tuning is to:

1. Shuffle the dataset order before each training epoch begins
2. Pad variable-length examples into a batched, masked tensor
3. Compute the loss over the batch after the forward pass
4. Convert raw text into token ids using the tokenizer

Answer: B — Pad variable-length examples into a batched, masked tensor

B) Correct — collation turns a list of per-example dicts into padded tensors with the matching attention mask and label mask. A) Shuffling is the sampler's responsibility. C) Loss is computed by the model or training loop. D) Tokenization normally happens in the dataset map step before collation.

A team exports a decoder LLM to ONNX and finds outputs diverge from PyTorch only for long prompts. The MOST likely cause is:

1. ONNX cannot represent attention or softmax operations at all
2. The graph was traced at fixed sequence length, not dynamic axes
3. The tokenizer was not embedded inside the exported ONNX graph file
4. ONNX always executes in INT8, so numeric precision is lost

Answer: B — The graph was traced at fixed sequence length, not dynamic axes

B) Correct — tracing without dynamic axes bakes in the example shape, so longer sequences are silently mishandled. A) Attention decomposes into standard ONNX operators. C) Tokenizers are normally external, which affects nothing about long-prompt numerics. D) ONNX preserves the exported precision; INT8 is an explicit choice.

A batch of prompts of very different lengths is passed to a decoder-only model for generation. To get correct results the developer must MOST importantly:

1. Use left padding and pass the attention mask
2. Sort the batch by descending length before tokenizing
3. Set truncation=True so every prompt is the same length
4. Disable the KV cache so padding tokens are not stored

Answer: A — Use left padding and pass the attention mask

A) Correct — decoder-only generation continues from the last position, so pads must sit on the left, and the mask keeps them out of attention. B) Length sorting is a throughput optimization and does not fix correctness. C) Truncation discards prompt content and still leaves length differences unhandled. D) Disabling the cache slows generation without addressing where the padding sits.

Which practice BEST supports experiment reproducibility during fine-tuning?

1. Record the random seed, data snapshot hash and full config per run
2. Save only the final checkpoint and the best validation score
3. Reuse one shared output directory for every experiment
4. Keep hyperparameters in a notebook cell edited between runs

Answer: A — Record the random seed, data snapshot hash and full config per run

A) Correct — seed, data version and config together let a run be recreated; any one alone is insufficient. B) A score without the recipe cannot be reproduced. C) A shared directory overwrites artifacts and destroys the audit trail. D) Ad-hoc notebook edits leave no record of what produced which result.

Which artifact should NOT be committed directly to a Git repository for an LLM application?

1. The requirements or lock file that pins dependency versions
2. Multi-gigabyte model checkpoint binaries used at inference
3. The Dockerfile that defines the runtime image
4. Unit tests covering the prompt-construction helpers

Answer: B — Multi-gigabyte model checkpoint binaries used at inference

B) Correct — large binaries bloat history irrecoverably; use a model registry or an LFS-style pointer instead. A) Lock files are small text and belong in the repo. C) Dockerfiles are source and must be versioned. D) Tests are source code and belong with the project.

A CI pipeline for an LLM application should run which check on every pull request WITHOUT calling the live model provider?

1. A full human evaluation of answer helpfulness and tone
2. A load test at production peak concurrency and payload size
3. A retraining run to confirm the loss curve still converges
4. Unit tests of prompt assembly against mocked responses

Answer: D — Unit tests of prompt assembly against mocked responses

D) Correct — deterministic logic around the model is exactly what mocks make fast, cheap and reliable in CI. A) Human evaluation cannot gate a per-PR pipeline. B) Load tests belong to a scheduled stage, not every commit. C) Retraining is far too slow and expensive for PR gating.

A developer wants the fastest path from a model name on the Hugging Face Hub to a working sentiment classifier in three lines of Python. The MOST appropriate Transformers API is:

1. The pipeline() factory, which bundles tokenizer and model
2. AutoModel.from_pretrained() followed by manual logit decoding
3. The Trainer class configured with an evaluation-only TrainingArguments
4. A raw torch.load() of the checkpoint plus a hand-written forward pass

Answer: A — The pipeline() factory, which bundles tokenizer and model

A) Correct — pipeline() is the high-level abstraction that wires tokenization, model inference and label mapping together for a task. B) Works but is the lower-level path: you must map logits to labels yourself, so it is not the fastest. C) Trainer is for training/evaluation loops over datasets, not ad-hoc inference. D) torch.load() bypasses the Hub config and tokenizer entirely and is the most manual option.

A LangGraph workflow must resume after a crash without replaying expensive tool calls. The feature that MOST directly enables this is:

1. Streaming callbacks attached to each node in the graph
2. A larger context window on the underlying chat model
3. Setting temperature to zero across all model nodes in the graph
4. State checkpointing persisted after each node transition

Answer: D — State checkpointing persisted after each node transition

D) Correct — durable checkpoints of graph state allow resumption from the last completed node. A) Callbacks observe execution but store no recoverable state. B) Context size does not persist anything across a process restart. C) Determinism helps testing but does not save progress.

Why are containers the standard delivery unit for GPU AI workloads, and what extra component is required over a normal container?

A container pins the framework, CUDA runtime, libraries, and Python dependencies into one immutable image, so the environment that passed testing is the environment that runs in production. GPU workloads additionally need the NVIDIA container toolkit and a compatible host driver to expose the devices — the driver stays on the host and is not inside the image.

Why must a tokenizer always be loaded from the same checkpoint as its model?

The model's embedding matrix is indexed by token IDs from a specific vocabulary. A mismatched tokenizer produces IDs that map to unrelated embeddings and special tokens the model never saw, so the output is fluent-looking nonsense rather than an obvious error — a failure that is easy to ship undetected.

What problem does LangGraph solve that a linear LangChain chain does not?

A linear chain runs a fixed forward sequence. LangGraph models the application as a directed graph of nodes with explicit shared state, so it supports cycles, conditional branching, retries, human-in-the-loop pauses, and checkpointed resumption — the control flow an agent needs when it must loop between planning, tool calls, and reflection.

What is the NVIDIA NGC catalog and what kinds of artifacts does it host?

NGC is NVIDIA's hub for GPU-optimized AI software. It hosts containers for frameworks and inference servers, pretrained models, Helm charts, and SDKs, all built and tested against specific CUDA and driver stacks so a pull-and-run gives a known-good environment.