SLDA always produce the same response variable estimate for all the documents

[zli12321]

I have a size 13000 dataset with 20 categories and trained SLDA with those labels with K=16. After training, I first call the infer function, then estimate to predict a document from the corpus. For every document trained for the model, the estimate result always returns 19 as the argmax category. Are there any bugs with supervised LDA code?