NCA-GENL Software Development Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Enable pipeline parallelism across the available GPUs to distribute the long prompts across multiple devices.

  Why it's wrong here

  Pipeline parallelism splits model layers across GPUs to fit large models, but it does not change the per-sequence token budget. A sequence that hits the maximum length limit will still truncate regardless of how many devices hold the layers. The symptom points to a length constraint rather than insufficient model capacity, so adding pipeline stages addresses the wrong bottleneck.

• ✗

  Lower the sampling temperature and top-p values so the model produces shorter, more deterministic completions.

  Why it's wrong here

  Sampling parameters influence which tokens are chosen, not when generation is forced to stop. Truncation at a fixed boundary with no error is a capacity limit symptom, not a distributional one. Adjusting temperature or top-p would change style and variability while leaving the mid-sentence cutoff intact, so it would not resolve the reported behavior under concurrent long prompts.

• ✓

  Increase the maximum sequence length setting so the combined prompt and generated tokens fit within the engine's configured limit.

  Why this is correct

  When the combined prompt plus generated tokens exceed the engine's maximum sequence length, generation stops at the boundary and returns a truncated response without raising an error. Long concurrent prompts consume that budget quickly, so raising the maximum sequence length gives generation room to finish. This directly explains silent truncation with no error logged and no crash.

• ✗

  Reduce the maximum batch size so each concurrent request receives a larger share of the KV cache blocks.

  Why it's wrong here

  Batch size governs how many sequences run together and how cache blocks are shared, but it does not extend the maximum sequence length. If the configured length cap is the stopping condition, shrinking the batch changes scheduling pressure without lifting the cap. Requests would still truncate at the same token boundary, so this adjustment misses the actual cause of the mid-sentence cutoffs.