3 ms·

Thanks! I'll try. I didn't bother believing that if this was developed heavily on CUDA, it was likely going to use kernels that were missing in MPS.