C++ failure while compiling cmdstanpy model with reduce_sum_static

Modeling
1

Hello all. My model with reduce_sum_static fials to compile with the error that begins with the following (the entire error is really long):

ROR:cmdstanpy:file model.stan, exception ERROR
 In file included from /include/c++/4.9.3/bits/move.h:57:0,
                 from /include/c++/4.9.3/bits/stl_pair.h:59,
                 from /include/c++/4.9.3/bits/stl_algobase.h:64,
                 from /include/c++/4.9.3/bits/char_traits.h:39,
                 from /include/c++/4.9.3/ios:40,
                 from /include/c++/4.9.3/istream:38,
                 from /include/c++/4.9.3/sstream:38,
                 from /include/c++/4.9.3/complex:45,
                 from stan/lib/stan_math/lib/eigen_3.3.7/Eigen/Core:96,
                 from stan/lib/stan_math/lib/eigen_3.3.7/Eigen/Dense:1,
                 from stan/lib/stan_math/stan/math/prim/fun/Eigen.hpp:22,
                 from stan/lib/stan_math/stan/math/rev.hpp:4,
                 from stan/lib/stan_math/stan/math.hpp:19,
                 from stan/src/stan/model/model_header.hpp:4,
                 from model.hpp:3:
/include/c++/4.9.3/type_traits:512:12: error: template instantiation depth exceeds maximum of 900 (use -ftemplate-depth= to increase the maximum) instantiating ?struct std::__or_<std::is_lvalue_reference<std::_Tuple_impl<63ul, const int&>&>, std::is_rvalue_reference<std::_Tuple_impl<63ul, const int&>&> >?
     struct is_reference
            ^

The model is attached. Any ideas what’s wrong will be highly highly appreciated!

model.txt (6.6 KB)

2

The model you supplied has a syntax error:

functions {

real partial_sum(real[] dummy, int start, int end, real[,] X, vector mu_prior_x, real sigma_v,
                                matrix A, matrix B, real[,,] U, real Q, matrix C, real sigma_r,

                                int[,,] pressed_gng, int[,,] cue_gng, real[,,] outcome_gng, int[,] Tr_gng, int[] idx_gng_obs, 
                                real[] b_gng_pr, real[] pi_gng_pr, real[] xi_gng_pr, real[] ep_gng_pr, real[] rho_gng_pr,
                                real[] b_gng, real[] pi_gng, real[] xi_gng, real[] ep_gng, real[] rho_gng, vector initV_gng, int Bl) {   

    real lt = 0;

    if (w == 1) {                                           
        lt += normal_lpdf(to_vector(X[start,])|mu_prior_x,sigma_v);                           
    }

w is not defined before that.

Also please specify the C++ compiler & OS.

3

So sorry, I attached the wrong model file. Even with w I get the same error.
OS is centOS7 and C++ compiler is 4.9.3

model.txt (6.6 KB)

4

Oh, so you only got the sysadmins to upgrade from 4.8 to 4.9.3? That is unfortunate as that is quite an old compiler (5years+). It should still work though.

The generated C++ from your Stan model is complex and g++ 4.9.3 is not the most efficient with it so its default settings are not able to handle compiling it. Let’s see if setting the limits higher would work:

Put
CXXFLAGS_PROGRAM += -ftemplate-depth=2000
in your make/local.

FTR the model compiles fine with default settings with g++ 9.3.0 and clang++.

5

@syclik @stevebronder @Bob_Carpenter @bbbales2 @wds15 @tadej

Just a heads up that we might really need to start thinking about setting the compiler bar a bit higher than g++ 4.9.3 or we will need to stop increasing compile-time complexity of our codebase.

We might get around it with overriding defaults for some more time, but not much longer.

6

Yup. That’s what major version bumps can be used for.

7

That’s super helpful! Is there a tutorial on how to make changes to make/local? Bc I saw I will have to edit make/local also to make MPI work.

Thanks!

8

The instructions are:

In your cmdstan folder (if you installed with cmdstanpy check the home folder for .cmdstanpy) you should find a folder named make. Create a file named local in it and put the above line in that file.

9

Linking to the previous discussion here

The tl;dr is that if we want to bump up we should get C++17 along the way. we could bump up to gcc-7* and leave clang* as is. But to do this we need to fixup the pystan wheel builder so that it uses the python package docker builder with gcc-8. I’m for having a 2-3 months deprecation notice before the bump (so about 9 month deprecation notice)

*clang-6 has c++17 support but there are some defect reports that weren’t resolved till later version. Same goes for gcc-7 where some defect reports weren’t resolved till gcc-8.

10

UPDATE - I updated to g++ 9.3.0 and the model compiled just fine, but I have a sampling error now:

RuntimeError: Error during sampling.
chain 1 returned error code -11

Should I recomplie cmdstan from scratch using g++ 9.3.0, or that’s fine using the cmdstan version I downloaded with g++ 4.9.3 loaded?

11

You definitely need to recompile cmdstan.

12

Update 2 - the model compiles and begins to sample, but now exists with :

   raise RuntimeError(err_msg)
RuntimeError: Error during sampling.
chain 1 returned error code 70

Same happens for another model that worked just fine with g++ 4.9.3.

13

Autofix – it failed bc of bad initial values, which I changed and it samples alright now. Thanks!

14

So is error code 70 get triggered because of bad initial values?

15

You can read more on this here:

But in my code it was solved after changing the initial values.