Comment by woctordho

It would be good if we can use formal verification to see to which extent the quantization will overflow in intermediate results. There are some widely-known annoying bugs that SageAttention (int8 quantized attention) works on some models but produces black images on other models because of overflow, and currently no one knows how to use it in training. There should be a better way to prevent this.