We’ve been finding all sorts of weird issues with the latest rstan unfortunately. One option to is to use the older version (which I think you said on GitHub that you’ve done), but another option is to use the new CmdStanR interface, which should play nicer with R (i.e., fewer crashes). To run VB with CmdStanR the syntax is:
# http://mc-stan.org/cmdstanr/reference/model-method-variational.html
model <- cmdstan_model(stan_file)
fit <- model$variational(data, ...)