NVIDIA NCA-AMAD exam · Official objectives · Original practice content · Last reviewed October 16, 2026
Free NCA-AMAD Model Customization and Fine-Tuning Concepts Practice Questions
The Model Customization and Fine-Tuning Concepts domain is 14% of the NVIDIA NCA-AMAD exam and covers prompt engineering, RAG, chunking and hybrid search, SFT, LoRA and PEFT, preference optimization, distillation and adapter serving. Review each answer and explanation as you practice.
The NCA-AMAD exam has 50-60 multiple-choice questions plus a hands-on lab in 1 hour. It costs $125, results are pass or fail, and the certification is valid for 2 years.
Data requirements, catastrophic forgetting and overfitting
Serving fine-tuned models with adapters or merging, and validating impact with offline experiments and A/B tests
How to study Model Customization and Fine-Tuning Concepts
Frequently changing knowledge that must be cited points to RAG; a consistent new behavior or format points to fine-tuning.
Exact identifiers such as part numbers need keyword (hybrid) search, because dense embeddings blur them.
LoRA trains small low-rank matrices on a frozen base model, which is why many adapters can share one deployment.
NCA-AMAD Model Customization and Fine-Tuning Concepts sample questions with answers
Model Customization and Fine-Tuning Concepts
Sample question 1
What is the main advantage of merging a LoRA adapter into the base model's weights before deployment?
It shrinks the base model's weights so they need less GPU memory
It avoids the extra inference latency of managing a separate adapter
It lets operators swap in a new adapter without rebuilding the model
It lets one deployment serve many different tasks in mixed batches
Show correct answer and explanation
Correct answer: B. It avoids the extra inference latency of managing a separate adapter
B is correct because merging adds the low-rank update into the pretrained weights, giving a model that is structurally the same as the base model, so no separate adapter has to be handled at inference time. A is incorrect because the merged weights have the same shape and size as the original weight matrices; merging adds no compression. C is incorrect because once the adapter is merged it is part of the weights; swapping adapters at runtime is what dynamic loading provides. D is incorrect because a merged model becomes bespoke and serves only the task it was tuned for; mixed-task batching is the benefit of loading adapters dynamically.
Model Customization and Fine-Tuning Concepts
Sample question 2
An IT service team has trained three LoRA adapters on the same base model: one for bug summarization, one for ticket routing and one for root-cause analysis. All three see steady traffic. Which serving setup uses the GPUs MOST efficiently?
One NIM with the base model that loads the right adapter for each request
Three merged models, each deployed on its own set of GPUs
One merged model that combines the weights of all three adapters
The base model only, with the required task named in the system prompt
Show correct answer and explanation
Correct answer: A. One NIM with the base model that loads the right adapter for each request
A is correct because dynamic LoRA serving keeps one copy of the base model and applies the requested adapter per request, so requests for different tasks can share batches without hosting separate custom models. B is incorrect because three merged models triple the base-model memory and cannot batch requests across tasks; merging is recommended only when a deployment serves a single task. C is incorrect because merging produces a model specialized for one task; adding three task updates into the same weights does not keep each task's behavior separate. D is incorrect because this discards the fine-tuning the team already did; the adapters exist because prompting the base model was not accurate enough.
Model Customization and Fine-Tuning Concepts
Sample question 3
A bank LoRA-tuned an 8B model to classify loan documents. It is the only task on a dedicated GPU deployment, the adapter is retrained about twice a year, and the latency target is tight. Which deployment approach is MOST appropriate?
Repeat the training as full fine-tuning so that no adapter needs to be served
Merge the adapter into the base weights and serve the merged model
Host the base model and the adapter on separate GPUs to spread the load
Serve the adapter dynamically from an adapter store alongside the base model
Show correct answer and explanation
Correct answer: B. Merge the adapter into the base weights and serve the merged model
B is correct because with one task per deployment and rare updates, merging gives a purpose-built model with no adapter-management overhead at inference, which suits the tight latency target. A is incorrect because full fine-tuning costs far more compute and is unnecessary; merging the existing adapter already yields a standalone model. C is incorrect because the adapter is a small set of low-rank matrices applied inside each layer; placing it on another GPU would add transfer overhead, not reduce latency. D is incorrect because dynamic loading adds some latency and throughput cost in exchange for flexibility across many adapters, which a single-task deployment does not need.
Model Customization and Fine-Tuning Concepts
Sample question 4
An engineer starts an LLM NIM with this command, and the directory ~/loras holds two adapters named math-v1 and support-v2:
How does a client application send a request that uses the support-v2 adapter?
It restarts the container with only support-v2 left in the directory
It passes the adapter's file path in a custom HTTP header on each call
It sets the model field in the request body to the adapter's name
It sends the request to a separate port that NIM opens for each adapter
Show correct answer and explanation
Correct answer: C. It sets the model field in the request body to the adapter's name
C is correct because NIM serves every adapter in the mounted PEFT directory under its own name, and the client picks one by naming it in the model field of an ordinary completions request. A is incorrect because NIM serves all adapters in the directory at once; restarting to change adapters would defeat dynamic loading. B is incorrect because clients refer to an adapter by its name in the request body, not by a file path; the mount already gives NIM the files. D is incorrect because the command publishes one port, 8000, and all adapters are reached through the same endpoint.
Model Customization and Fine-Tuning Concepts
Sample question 5
When an LLM NIM receives a request that names one of its LoRA adapters, how does it make that adapter available for the request?
It downloads a separate full model that was built for that adapter
It merges the adapter into the base weights and then restarts the service
It rebuilds the inference engine with the adapter compiled into it
It pulls the adapter from the adapter store into a GPU and host memory cache
Show correct answer and explanation
Correct answer: D. It pulls the adapter from the adapter store into a GPU and host memory cache
D is correct because NIM loads the named adapter from its adapter store into a multi-tier cache, keeping recently used adapters in GPU memory and others in host memory. A is incorrect because each NIM is tied to a single foundation model; adapters are small add-ons to it, not separate full models. B is incorrect because NIM keeps adapters separate from the base model so that many adapters can be served without restarts. C is incorrect because adapters are loaded at request time; rebuilding an engine per adapter would make dynamic serving impractical.
Model Customization and Fine-Tuning Concepts
Sample question 6
An LLM NIM serves the Llama 3 8B foundation model. Which LoRA adapters can this NIM load and serve?
Only adapters that were trained with NVIDIA NeMo Customizer
Only adapters that were first merged into full model weights
Any adapter trained on a model of the same family, such as Llama 3 70B
Adapters trained on Llama 3 8B itself, the NIM's base model
Show correct answer and explanation
Correct answer: D. Adapters trained on Llama 3 8B itself, the NIM's base model
D is correct because each NIM is associated with a single foundation model, and an adapter holds low-rank updates to that model's weights, so only adapters trained on the same base model apply. A is incorrect because NIM accepts adapters trained with either the NVIDIA NeMo framework or the Hugging Face PEFT library. B is incorrect because a merged adapter becomes a full standalone model; NIM serves separate adapters dynamically on top of its base model. C is incorrect because an adapter's low-rank matrices match the layer shapes and weights of the exact model it was trained on, and a 70B model has different layers.
Model Customization and Fine-Tuning Concepts
Sample question 7
To keep more customer adapters resident, an operator raises an LLM NIM's settings for the maximum number of loaded adapters and the maximum adapter rank. The same profile ran fine before, but the service now fails at startup with GPU out-of-memory errors. What is the BEST fix?
Raise the rank further so that fewer adapters are needed for each customer
Go back to supported values, or pick a profile with more memory
Restart the service so that the adapter cache starts out empty
Merge all of the customer adapters into the base weights to save memory
Show correct answer and explanation
Correct answer: B. Go back to supported values, or pick a profile with more memory
B is correct because adapter weights need memory in addition to the base model, and raising the adapter count or maximum rank can exhaust it; the fix is supported settings or a profile with enough memory for the required adapters. A is incorrect because a higher rank makes each adapter larger and would increase, not decrease, the memory reserved for adapters. C is incorrect because the failure comes from the configured adapter limits, not from adapters left over in a cache, so a restart with the same settings fails the same way. D is incorrect because a merged model serves only one task, so merging many customers' adapters into one model would not preserve each customer's behavior.
Model Customization and Fine-Tuning Concepts
Sample question 8
A platform team serves 40 LoRA adapters on one base model. Data scientists trained them with ranks from 8 to 128. Why might the operators ask that future adapters all use one fixed rank?
A fixed rank removes the need to evaluate each adapter before it is deployed
A fixed rank lets every adapter be merged into the base weights at the same time
Adapters of uniform size batch together better, which improves serving performance
Adapters with mixed ranks cannot be loaded by NIM in the same deployment
Show correct answer and explanation
Correct answer: C. Adapters of uniform size batch together better, which improves serving performance
C is correct because the NIM team notes that operators may enforce a uniform adapter size because uniform LoRAs enable better batching and therefore better performance. A is incorrect because each adapter still needs task evaluation; its rank says nothing about whether it is accurate. B is incorrect because merging still yields a single-task model regardless of rank; rank uniformity is about batching, not merging. D is incorrect because mixed-rank adapters can be served; they simply cost more latency and throughput than adapters of one rank.
Model Customization and Fine-Tuning Concepts
Sample question 9
A team benchmarks one base model three ways: base model alone, a single merged LoRA model, and dynamic serving of 30 LoRA adapters with mixed ranks. Dynamic serving shows higher latency and lower throughput than the other two. Which conclusion is MOST accurate?
The overhead is expected and is the price of serving many tasks from one deployment
The benchmark is faulty, because dynamic adapters should add no measurable overhead
Each adapter loads its own full copy of the base model, which slows every request
Merging all 30 adapters into one model would keep each task's tuned behavior
Show correct answer and explanation
Correct answer: A. The overhead is expected and is the price of serving many tasks from one deployment
A is correct because compared with a base or merged model, dynamic LoRAs, especially many adapters of different ranks, add latency and throughput cost; the team should judge whether that cost is reasonable for the flexibility gained. B is incorrect because NVIDIA's guidance states that single, multiple and mixed-rank dynamic LoRAs all add increasing cost compared with the base model. C is incorrect because all adapters share one base model; only the small low-rank matrices are loaded per adapter. D is incorrect because a merged model is bespoke to one task, so it cannot reproduce 30 different adapters' behavior.
Model Customization and Fine-Tuning Concepts
Sample question 10
After a new model passes offline evaluation, what is the usual way to measure its effect on real users before a full rollout?
An A/B test that sends part of the live traffic to the new model
Having the new model rate the quality of its own production responses
Running the same offline test set again every day for several weeks
Comparing the final training loss with the current model's training loss
Show correct answer and explanation
Correct answer: A. An A/B test that sends part of the live traffic to the new model
A is correct because after training and offline evaluation, models typically move to online evaluation through A/B testing, comparing the new model with the current one on real traffic. B is incorrect because self-rating is biased and does not compare the new model with the current one on the same traffic. C is incorrect because repeating the offline set measures nothing new about real users; online behavior needs live traffic. D is incorrect because training loss shows fit to training data, not the effect on users or business metrics.
FlashGenius is an independent study resource and is not affiliated with or endorsed by NVIDIA. NVIDIA and NCA-AMAD are trademarks of NVIDIA Corporation. 600 original practice questions.