unconstrain_pars() crashes R

Summary:

unconstrain_pars causes segfault in R

Description:

When running unconstrain_pars on a newly fitted model, it causes a segfault and crashes R.

Reproducible Steps:

Standard eight schools model

data {
  int<lower=0> J;                                                                                                   
  real y[J];                                                                         
  real<lower=0> sigma[J];                                    
}
parameters {
  vector[J] theta_trans;                                                                                       
  real mu;                                                                                            
  real<lower=0> tau;                                                                                                
}
transformed parameters{
  vector[J] theta;                                                                                                             
  theta = theta_trans * tau + mu;
}
model {
  // priors                                                                                                                                       
  target += normal_lpdf(theta_trans | 0, 1);
  target += normal_lpdf(mu | 0, 5);
  target += cauchy_lpdf(tau | 0, 5);
  
  //likelihood                                                                                                                                    
  target += normal_lpdf(y | theta, sigma);
}

library(rstan)
model <- stan_model("eight_schools.stan")
d <- list(
  J = 8,
  y = c(28, 8, -3, 7, -1, 1, 18, 12),
  sigma = c(15, 10, 16, 11, 9, 11, 10, 18)
)
fit <- sampling(model, data = d)

unconstrain_pars(fit, pars = list(mu = 0, theta = rep(1.5, 8), tau = 2))

Current Output:

 *** caught segfault ***
 *** caught segfault ***
address (nil), cause 'memory not mapped'

Traceback:
 1: .External(list(name = "CppMethod__invoke_notvoid", address = <pointer:0x55a5f4813370>,     dll = list(name = "Rcpp", path =                  
"/lib/R/library/Rcpp/libs/Rcpp.so",         dynamicLookup = TRUE, handle = <pointer: 0x55a5f2b78520>, info = <pointer: 0x55a5f202dad0>),
numParameters = -1L),     <pointer: 0x55a5fb4cd110>, <pointer: 0x55a5f76d9760>, .pointer,     ...)
 2: object@.MISC$stan_fit_instance$unconstrain_pars(pars)
 3: .local(object, ...)
 4: unconstrain_pars(fit, pars = list(mu = 0, theta = rep(1.5, 8),     tau = 2))
 5: unconstrain_pars(fit, pars = list(mu = 0, theta = rep(1.5, 8),     tau = 2))

Expected Output:

Interestingly, the following works:

fit1 <- stan_demo("eight_schools")
unconstrain_pars(fit1, pars = list(mu = 0, theta = rep(1.5, 8), tau = 2))

[1] 0.0000000 1.5000000 1.5000000 1.5000000 1.5000000 1.5000000 1.5000000
 [8] 1.5000000 1.5000000 0.6931472

RStan Version:

2.26.1

R Version:

"R version 4.0.5 (2021-03-31)"

Operating System:

Ubuntu 18.04

Start with the RStan unconstrain_pars entry point shown in the traceback and compare its behavior with the working stan_demo("eight_schools") call. Reproduce the segfault using the supplied eight schools model, fitted object, and pars list under RStan 2.26.1, then verify that unconstrain_pars returns the expected transformed values without crashing R.