I have made multiple PRs to PyMC3, which were bug fixes, documentation and small feature additions.
I found a bug in PyMC3's multinomial random variate sampler, related to floating point precision issues while moving numbers from the GPU to the CPU, when working on my Bayesian analysis recipes repository. Specifically, we get probability values that sum to infinitesimally larger than one. I thus submitted a patch that fixed that converts the probabilities to
float64 precision and re-normalizes the probabilities to 1, before using them for random sampling.
I contributed small changes to the docs, to make it super clear that precision and standard deviation parameterizations were alternate parameterizations, and only one was needed.
I contributed the definition of a "mode" for the Weibull distribution, so that I could use it in a Mixture Weibull.