quantization-aware training (qat)

Quantization-Aware Training (QAT) trains models with quantization effects simulated, yielding better low-precision accuracy than PTQ. **Mechanism**: Insert fake quantization nodes during training, forward pass simulates quantized behavior, gradients computed through straight-through estimator (STE), model learns to be robust to quantization noise. **Why better than PTQ**: Model adapts weights to quantization-friendly distributions, learns to avoid outlier activations, can recover accuracy lost in PTQ especially at very low precision (INT4, INT2). **Training process**: Start from pretrained FP model, add quantization simulation, fine-tune for additional epochs, export quantized model. **Computational cost**: 2-3x training overhead due to quantization simulation, requires representative training data, more complex training pipeline. **When to use**: Target precision is INT4 or lower, PTQ results unacceptable, have training infrastructure and data, accuracy is critical. **Tools**: PyTorch FX quantization, TensorFlow Model Optimization Toolkit, Brevitas. **Trade-offs**: Better accuracy than PTQ but requires training, best when combined with other compression techniques (pruning, distillation).