SDPA Int8 Quantisation using MTQ

[satya-penamakuri]

I'm attempting to quantize SDPA similarly to TRT's approach—by adding Q-DQ on the Q, K, V, and softmax output. I've observed that MOQ implements this, whereas MTQ does not. Additionally, I would like to perform QAT after adding quantizers via MTQ. Does MTQ offer kernels to mitigate training slowdowns? This is crucial since SDPA typically features efficient implementations, and the quantization process requires inserting a Q-DQ at an intermediate output.