Does cvxpylayer support Gurobi as solver, and then back-propagate?

[keungRX]

Does cvxpylayer support Gurobi as solver, and then back-propagate? I am trying to solve a large_scale MIP problem in cvxpy layer in my neural network, and SCS is incapable to solve my problem.

[akshayka]

No it does not.

[holmesco]

I implemented this on my fork of cvxpylayers so that I could use Mosek.
You can use any solver as long as it gives you access to the primal, dual and slack variables. It basically injects the primal-dual solution on the forward pass so that it is used in the implicit differentiation on the backward pass. Note that it also requires my fork of the diffcp module, which does all the heavy lifting for cvxpylayers. Also I only implemented the PyTorch version.