`normal_rng` crashes when provided zero-length vector for mu and scalar for sigma

[rmcminds]

Environment:

• CmdStan 2.37.0
• cmdstanr 0.9.0
• Windows 11 (Windows_NT 10.0.26200), x64, g++ toolchain from RTools

Calling normal_rng with a zero-length vector mu and scalar sigma in generated quantities crashes the CmdStan executable with an access violation (Windows exit code -1073741819 / 0xC0000005) and no error message. Via cmdstanr the only symptom is Warning: Chain 1 finished unexpectedly!.

This was very difficult to debug given the lack of an informative error message. I went down the wrong rabbit hole for a while because debugging print statements worked in transformed data but didn't print in transformed parameters; even though the error wound up being in what I figured would be executed after transformed parameters (generated quantities).

I fixed my code by using a logical gate, but it seems like that shouldn't be necessary - it return a zero-length result or reject() with an informative message.

R reprex (cmdstanr)

library(cmdstanr)

stan_file <- write_stan_file('
data {
  int<lower=0> N;
}
transformed data {
  print("TD");
}
parameters {
  real mu;
}
transformed parameters {
    print("TP");
}
model {
  mu ~ std_normal();
}
generated quantities {
  real marker = 0;
  vector[N] z;
  print("GQ");
  z = to_vector(normal_rng(zeros_vector(N), 2.5));
  marker = 1;
}
')

mod <- cmdstan_model(stan_file)

fit <- mod$sample(data = list(N = 1), chains = 1, iter_warmup = 1, iter_sampling = 1)

# Works:
gq_ok <- mod$generate_quantities(fitted_params = fit, data = list(N = 1))

# Crashes silently after the TD block but before TP:
gq_bad <- mod$generate_quantities(fitted_params = fit, data = list(N = 0))

# Sampling with N = 0 also crashes (at the first post-warmup GQ evaluation):
fit_bad <- mod$sample(data = list(N = 0), chains = 1, iter_warmup = 1, iter_sampling = 1)

[rmcminds]

Also I just updated to cmdstan 2.39.0, and have the same issue. Should have done that before!

[WardBrian]

This seems like it could be an issue with many of our RNGs, since none of them seem to call size_zero like the lpdfs do