Is there a way to activate int8 MHA_v2 kernel when SeqLen > 512?

Description

Hi, I notice form Issue that the int8 MHA_v2 kernel only supports SeqLen <= 512. I also try on my own diffusion model whose Q shape is (B, N, S, H) and S >> 512. I use pytorch_quantization to insert QDQ in the MHA and convert to TRT. As expected, it breaks into 3 kernels( int8->gemm->FP32->softmax->int8->gemm->int8) and runs slower than fp16 MHA_v2.

But I notice the TensorRT 9.3 oss introduces ammo to quantize SDXL, and by veiwing the code I assume the MHA is also quantized because there are codes dealing with the QKV QDQ fusion. So dose the demo SDXL manage to involke int8 MHA_v2 kernel? I think the SDXL SeqLen is also >> 512. How did the demo quantization manage to utilize int8 MHA?