Advanced Topic: Using External C++ Functions

This is based on the relevant portion of the CmdStan documentation here

Consider the following Stan model, based on the bernoulli example.

from cmdstanpy import CmdStanModel
model_external = CmdStanModel(stan_file='bernoulli_external.stan', compile=False)
functions {
  real make_odds(real theta);
data {
  int<lower=0> N;
  array[N] int<lower=0, upper=1> y;
parameters {
  real<lower=0, upper=1> theta;
model {
  theta ~ beta(1, 1); // uniform prior on interval 0, 1
  y ~ bernoulli(theta);
generated quantities {
  real odds;
  odds = make_odds(theta);

As you can see, it features a function declaration for make_odds, but no definition. If we try to compile this, we will get an error.

INFO:cmdstanpy:compiling stan file /home/docs/checkouts/ to exe file /home/docs/checkouts/
ERROR:cmdstanpy:Stan program failed to compile:
--- Translating Stan model to C++ code ---
bin/stanc  --o=/home/docs/checkouts/ /home/docs/checkouts/
Semantic error in '/home/docs/checkouts/', line 2, column 2 to column 29:
     1:  functions {
     2:    real make_odds(real theta);
     3:  }
     4:  data {

Some function is declared without specifying a definition.
make: *** [make/program:50: /home/docs/checkouts/] Error 1

Command ['make', '/home/docs/checkouts/']
        error during processing No such file or directory

Even enabling the --allow-undefined flag to stanc3 will not allow this model to be compiled quite yet.

INFO:cmdstanpy:compiling stan file /home/docs/checkouts/ to exe file /home/docs/checkouts/

To resolve this, we need to both tell the Stan compiler an undefined function is okay and let C++ know what it should be.

We can provide a definition in a C++ header file by using the user_header argument to either the CmdStanModel constructor or the compile method.

This will enables the allow-undefined flag automatically.

INFO:cmdstanpy:compiling stan file /home/docs/checkouts/ to exe file /home/docs/checkouts/
INFO:cmdstanpy:compiled model executable: /home/docs/checkouts/

We can then run this model and inspect the output

fit = model_external.sample(data={'N':10, 'y':[0,1,0,0,0,0,0,0,0,1]})
INFO:cmdstanpy:CmdStan start processing

INFO:cmdstanpy:CmdStan done processing.

array([0.361904, 0.954058, 0.420802, ..., 0.581759, 0.42715 , 0.266583])

The contents of this header file are a bit complicated unless you are familiar with the C++ internals of Stan, so they are presented without comment:

#include <boost/math/tools/promotion.hpp>
#include <ostream>

namespace bernoulli_model_namespace {
    template <typename T0__>  inline  typename
          make_odds(const T0__& theta, std::ostream* pstream__) {
            return theta / (1 - theta);