Reproducibility of the 1900-unit model

I found that when running the 1900-unit model with the same protein sequence input, it produces slightly different embedding: 82% of the difference occur on the 10^-8 to 10^-7 range. 64-unit and 256-unit produce identical results. 

Could you let me know how to fix this issue? Is it caused by floating-point arithmetic? It is to my understanding that the resulted embedding should be identical. 