Why fmincon's interior-point algorithm gives infeasible points for feasibility problem?

I am creating an initial feasible population for my optimization problem along the lines of this. I tried the technique with different solvers included in fmincon and noticed that albeit slowly, the medium-scale solvers sqp and active-set could give me feasible points but some points returned by the large-scale interior-point did not satisfy the constraints. Why this happens?

My problem has about 2000 linear inequality constraints and about 1500 variables, where none of them are integers.

[Sean de Wolski]

Post the MAT file containing the constraints and the code you use to optimize. IP should not be returning infeasible points and it is likely a bug in your constraints.