Michele LaurelliBlog

Quantization-aware training

Quantization-aware training: Training that simulates quantization effects to make the model more robust to reduced precision. Fake-quant operations introduce errors during learning, allowing parameters to adapt. It requires more work than post-training quantization but can improve low-bit quality. Examples include Fake quantization, INT8 training, Quality recovery at 4 bits.

In italiano: Quantization-Aware TrainingLegge IV · Loss, gradiente e ottimizzatori

Fake-quant operations introduce errors during learning, allowing parameters to adapt. It requires more work than post-training quantization but can improve low-bit quality.

Examples

  • Fake quantization
  • INT8 training
  • Quality recovery at 4 bits