LBFGS_FLOAT=32 results in error "LBFGSERR_ROUNDING_ERROR"

[markusdr]

I tried to use LBFGS_FLOAT=32 (i.e., float instead of double), to save memory in a large optimization task.

But it always aborts with LBFGSERR_ROUNDING_ERROR. I've tried this on different optimization problems. This error happens in later iterations, closer to convergence. Is there any way to prevent that error? I tried increasing xtol, but it doesn't help.

One way would be to fiddle with convergence criteria to declare it converged earlier, but I'd rather not do that.