Quantization-Aware Training

Quantization-Aware Training (QAT) is a torchao workflow/API for applying fake quantization during model training or fine-tuning so the converted quantized model can retain better accuracy or perplexity than post-training quantization. In torchao it uses prepare and convert steps, including QATConfig and quantize_ APIs, to insert fake-quantized layers before training and convert them to quantized operations afterward.