Calibration
Running representative inputs through a model to estimate the ranges or scales a quantized deployment will need.
Calibration runs representative inputs through a model before or during quantization so the system can choose realistic ranges and scales, because a low-bit deployment usually fails when those settings are learned from traffic that does not resemble the real job.
What It Is
Calibration is a short measurement pass over representative inputs. The model is observed in a higher-precision mode so the quantization pipeline can estimate ranges, scales, zero-points, or clipping thresholds for later low-bit execution. The exact math differs by method, but the practical goal is simple: choose numeric boundaries that fit normal deployment traffic.Why Data Quality Matters
Calibration works only if the sample inputs resemble the prompts, sequence lengths, and activation patterns the model will actually see in production. A tiny sample can still be useful, but it has to cover the right kinds of requests. When the calibration set misses important outliers or long-context behavior, the chosen ranges can be too tight or too loose.What Can Go Wrong
Bad calibration data can push quality down in several ways. Saturated activations lose detail because they no longer fit the chosen range; overly wide ranges waste the available low-bit precision; and workload shifts after deployment can expose layers that looked safe during calibration but were not truly representative. That is why calibration sensitivity often shows up first on activation-heavy quantization paths.Compared To Nearby Workflows
Calibration is usually one step inside a post-training quantization workflow, not a separate model family. It matters most when quantized activations or other runtime tensors need realistic ranges. Weight-only methods can sometimes need less calibration pressure, while activation quantization often needs it much more.References
- Jacob, Benoit, et al. "Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference." arXiv, 2017, https://arxiv.org/abs/1712.05877.
- Xiao, Guangxuan, et al. "SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models." arXiv, 2022, https://arxiv.org/abs/2211.10438.