Louis Bethune

I did not turn on `float64` on my initial test, [check by yourself!](https://colab.research.google.com/drive/1_lEj9d7juh6N-PRTLjmAn-sxvi2UEfEV?usp=sharing) I tested in `float32` in CPU / GPU / TPU in Colab; it works. In `float64` enabled,...

Thanks for the heads up. @jewillco: could you clarify your intent with this code? If my understanding is correct, you need: - a TPU for performance - `float64` precision enabled...

Hi! Indeed, `BoxOSQP` formulation does not make sense without equality constraints. I recall below the formulation solved by OSQP: $$\min_x \frac{1}{2}x^TQx+c^Tx\text{ s.t }l\leq Ax\leq u.$$ I recall the API: -...

> Also, does the problem formulation with an arbitrary mean that this can handle an arbitrary set of linear constraints, not just a box? Yes, it is the most general...

> Does that mean this class is a full QP solver, and not just for box-constrained problems? Yes, it is! Just like the original's one: https://osqp.org/

> It would also be nice to be able to provide only the initial solution point but not the dual information You can do this, as follows: ```python initial_solution_point =...

> The primal field needs to have two entries, and it looks like the duals are required as well or else the implementation crashes. It That's true! The method `init_params`...

Yes, [it is part of the doc](https://jaxopt.github.io/stable/_autosummary/jaxopt.BoxOSQP.html#jaxopt.BoxOSQP.init_params).

For consistency reasons of the API we are forced to take `params` KKT tuple in `run` and `update`, we cannot support an additionnal `init_x` without breaking implicit differentiation. I'm afraid...

Dear authors @IssamLaradji I am still interested in this topic and I would be glad if you could answer my concern. Thank you