NCP-GENL Fine-Tuning Practice Question
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?
⚠ Common exam trap
The trap here is assuming that a frozen base model or a low learning rate explains non-updating adapters, when the more likely cause is that the adapters are not attached to any executed module.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
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 analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
The base model is loaded in 8-bit precision, which prevents any gradient computation on the adapters.
Why it's wrong here
Quantized base models are commonly used with LoRA, and gradients still flow to the adapters because the adapters remain in higher precision. Loading the base in 8-bit does not block adapter training. If the adapters themselves were also quantized incorrectly, that could cause trouble, but the scenario does not indicate that.
- ✗
The learning rate is too low, so the adapter weights change by amounts below floating-point precision.
Why it's wrong here
An extremely low learning rate can slow learning, but it would still produce small nonzero updates and a gradually changing loss. A completely flat loss with no adapter movement over hundreds of steps is more consistent with adapters that are not connected to the computation graph at all, rather than merely updating very slowly.
- ✗
The LoRA rank is set too high, causing the adapters to be initialized to zero.
Why it's wrong here
A higher rank increases the number of trainable parameters but does not zero out the adapters. LoRA initialization already sets one of the low-rank matrices to zeros and the other to random values, so gradients can flow. A high rank would increase memory use and risk overfitting, but it would not produce a completely flat loss with no adapter updates.
- ✓
The target modules list is empty or points to modules that are not part of the model's forward pass.
Why this is correct
LoRA only updates the adapter matrices attached to the specified target modules. If the target modules list is empty or names modules that never execute, no adapter parameters participate in the forward pass, so no gradients reach them and the adapters remain unchanged. This exactly matches the symptom of frozen adapters and flat loss.
About these practice questions
This NCP-GENL question is part of Courseiva's 352-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCP-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCP-GENL exam.