NCP-AAI practice question 23 of 25

Which technique is MOST useful for decreasing response latency without retraining the base model?

Full set: 118 questions
Answer options
A: Inference optimization such as quantization or serving improvements Correct answer
B: Increasing prompt length for every request
C: Adding more manual review steps
D: Removing batching from the system
Correct Answer:
A. Inference optimization such as quantization or serving improvements
Explanation

To decrease response latency without retraining the base model, the most effective techniques are inference optimizations. These methods focus on making the model execute faster on the hardware. Quantization reduces the precision of the model's weights (e.g., from FP32 to INT8), which decreases memory footprint and speeds up computation. Serving improvements, such as using NVIDIA Triton Inference Server with a TensorRT-LLM backend, compile the model into a highly optimized engine for the target GPU architecture. These approaches directly reduce the time it takes to generate a response for a given input, addressing the latency requirement without altering the model's learned parameters.

Why Incorrect

B. Increasing prompt length adds more tokens for the model to process, which directly increases, not decreases, latency.
C. Adding manual review steps introduces human-in-the-loop delays, significantly increasing the total time to get a response.
D. Removing batching can hurt throughput and may not decrease single-request latency; optimized batching often improves overall system performance.

References

1. NVIDIA. (2024). NVIDIA TensorRT-LLM Documentation. Section: "Features". The documentation explicitly states that TensorRT-LLM provides "in-flight batching" and "INT8 and FP8 quantization" to optimize inference performance and reduce latency for large language models.

2. NVIDIA. (2024). NVIDIA Triton Inference Server Documentation. Section: "Optimization". This guide details how Triton can be configured for performance, including using optimized backends like TensorRT to "minimize latency."

3. Dettmers, T., et al. (2022). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. This academic paper demonstrates how 8-bit quantization can accelerate inference for large transformer models without significant performance degradation. DOI: https://doi.org/10.48550/arXiv.2208.07339

Scroll to Top