Currently, the unit tests for all the encoders are stochastic as they depend on gpt_sample which is random via the Mersenne twister. To make tests reproducible, we sample by taking the argmax of the logits.
Note that we can't fix the seed for both Python bark and bark.cpp.
Currently, the unit tests for all the encoders are stochastic as they depend on
gpt_sample
which is random via the Mersenne twister. To make tests reproducible, we sample by taking the argmax of the logits.Note that we can't fix the seed for both Python bark and bark.cpp.