Quantization
Domain: MLOps
🎯 Motivation
Reducing precision: 32-bit floats to 8-bit integers. Smaller model, faster inference, slight accuracy loss.
📋 Concrete Example
Quantization: 32-bit weights -> 8-bit weights. Model size: 1GB -> 250MB. Speed: 2x faster. Accuracy: 88% -> 86%. Worth the trade-off for edge deployment.
🔗 Analogy
Rounding numbers - 3.14159 -> 3.14. Slightly less precise but much easier to work with. Quantization rounds model weights for efficiency.