[QST] Column Generation with NVIDIA CuOPT

[shariharan2002]

Hello,
So I am trying to solve a very large LP in every iteration of column generation. This has approximately 20 million rows worth of constraints. I have tried some other commercial solvers with Barrier and Simplex - but they are seeming to take anywhere between 30 - 200 seconds per LP solve to prove optimality.
So, I was thinking of giving CuOPT a try. I do have access to an 80GB H100 GPU as well as about 8 cores and 8 X 16GB RAM.
But, I am worried about a few things:

1. So I need to retrieve the duals in every iteration - any existing example to do that?
2. I use Python , so an existing example of Column Generation using CuOPT would be great. Until now, for testing with CuOPT i just used to write the .mps file and then run it using cuopt_cli but now that is no longer good.
3. In every iteration I shall be adding columns -- is that bad? Will it lead to bottlenecks between transfer of data from CPU to GPU? Not sure

Thanks

[mlubin]

So I need to retrieve the duals in every iteration - any existing example to do that?

It's not clear to me what is meant by every iteration. You can access the duals after a completed solve in the documented Python API: https://docs.nvidia.com/cuopt/user-guide/latest/cuopt-python/convex/convex-api.html. You cannot access duals from intermediate iterations during a solve.

I use Python , so an existing example of Column Generation using CuOPT would be great.

We don't currently have an example of this. There's no special API for column generation, you would just follow the regular python API.

In every iteration I shall be adding columns -- is that bad? Will it lead to bottlenecks between transfer of data from CPU to GPU?

Yes, it will certainly be inefficient. cuOpt doesn't have a special code path for incrementally adding columns, so you'll have to transfer the entire problem into cuOpt for each solve.