liveBook · Manning

Calibration coverage

As explained in Quantization and Fast Inference

A comparison between the activation range estimated during calibration and the corresponding range observed in production.

Calibration coverage compares each layer’s calibration activation maximum with its production maximum: calibration max divided by production max, expressed as a percentage. Values below 100% indicate production activations exceed the calibrated range and may be clipped during quantized inference. section 4.2

Interpret the per-layer status thresholds as follows:

  • At or above 100%: OK; below 100%: WARNING; below 90%: CRITICAL. section 4.2

Layer-wise validation

A layer with calibration max 42.8 and production max 51.3 has 83.4% coverage and is critical. Raising its calibration max to 55 gives 107.2% coverage. section 4.2

How Quantization and Fast Inference covers it

How deep it goes
Defines a per-layer ratio and status thresholds, then shows validation and recalibration; it does not detail the script implementation.
Examples
Uses layer_6 at 83.4% coverage, then 107.2% after expanding its calibration range.
What it assumes
Requires per-layer calibration ranges and production activation statistics, gathered through instrumentation or a shadow run.
What you can do after
Diagnose undercovered layers by comparing calibration observer ranges with production activation logs.
How it connects
Connects calibration data and activation ranges to quantization and clipping.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Quantization and Fast Inference. The section links point to the parts of the book it draws on.