LESSON 6.3 · Theory · 120 min

Quantization: numbers, error, calibration

Quantization: numbers, error, calibration. Learn INT8/INT4, weight-only methods, GPTQ, AWQ, and GGUF and complete: Compare quality and speed across four quantization methods. Part

Learning objectives

  1. Explain what problem “Quantization: numbers, error, calibration” solves without hiding behind terminology.
  2. Trace the variables and causal links across INT8/INT4, weight-only methods, GPTQ.
  3. Complete “Compare quality and speed across four quantization methods” and judge the result with evidence rather than intuition.

Core concepts

INT8/INT4

INT8/INT4 is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment before moving to an optimized implementation.

weight-only methods

weight-only methods is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.

GPTQ

GPTQ is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.

AWQ

AWQ is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.

and GGUF

and GGUF is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.

Build and verify

Compare quality and speed across four quantization methods

  • Predict: write the expected output, trend, or failure before running code.
  • Build: implement only the minimum components needed to answer the question.
  • Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
  • Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.

Open the complete interactive lesson