With the help of our users we have identified a few issues with the latest release which is why we have just released CmdStan 2.26.1. The issues that were identified and fixed for this patch release are given below.
We would like to thank Github users cdriveraus, grburgess, rmcminds and spinkney for reporting the issues and our developers andrjohns, bbbales2, nhuurre and t4c1 for helping track down and fix the issues.
The Stan function cholesky_decompose not propagating derivatives correctly
This issue could have affected all models using cholesky_decompose. If you have used
CmdStan 2.26.0 with a model that uses cholesky_decompose we strongly advise re-runing the model.
Improper handling of input Eigen expressions with reduce_sum
This error caused some models that used reduce_sum to not compile from the generated C++ to the executable. The underlying issue was improper handling of Eigen expressions. There is no need to re-run your models if you were using reduce_sum with CmdStan 2.26.0.
Not allowing user defined _lp with a void return type in transformed parameters
A user-defined _lp function with a void return type, for example:
void lprob_lp(real x) {
target += normal_lpdf(x | 0, 1);
}
was not allowed in the transformed parameters block. User-defined _lp functions returning non-void were not affected.
Issue with row_vector inputs in distributions when OpenCL was enabled
Models using distributions with row_vector inputs failed to compile from C++ to executable with OpenCL enabled.
Models not using the OpenCL backend were not affected.
In addition to fixing the issues we have enhanced our testing to catch similar errors in the future before they are released. We apologize for any inconvenience or problems these may have caused.
The Stan Blog 