Restricted Boltzmann machine training

With the new pytorch integration option for restricted Boltzmann machines trained through Leap, I was trying to do some diligence about the strengths of the convention, as it appears that the NIPS research community had channeled a fair bit of energy into that space prior to the advent of GPU trained deep ReLU neural networks. It appears to be conventional wisdom that RBMs are most obviously suitable for modeling distributions of discrete variables, and for integration into a setting of continuous variables methods like insertion between linear activation layers a way to bridge, or G Hinton also has publications referring to conventions like “Gaussian–Bernoulli RBM” as well as attempts to integrate ReLU activations directly into an RBM setting. It appears part of the reason these methods are necessary is that if we were to attempt to extend the RBM variables to accommodate bitwise representations of continuous variables in the more traditional floating point arithmetic manner, the joint distributions have been considered ‘untrainable’ in prior literature. I am wondering if the D-Wave conventions for restricted Boltzmann machine training, and particularly those involving integration into a PyTorch setting, have attempted to validate whether continuous variables may be accommodated in a RBM setting with Advantage2 hardware? (Or are your hybrid solver systems already capable of the more exotic variable types in an RBM setting?) Cheers