Quantized training doesn't work. Quantization-aware training does, and that is when you use floats but adjust the math to compensate for the fact that you are going to quantize at inference time.