rasa: training is too slow

Question

I have rtx 3090 gpu and i9 12th gen processor. my training is not too large as well and yet the training time is too long. When I begin the training phase it says 24 cores available but limiting to safe limit of only 8 cores. NUMEXPR_MAX_THREADS not set.

Answer 1

In your terminal add the NUMEXPR_MAX_THREADS to your terminal. You can do so by writing in your CLI: export NUMEXPR_MAX_THREADS="24" if you want to use all of them. This will work until you close your terminal. You can add it permanently to your terminal profile (.bash_profile, ~/.zshrc ...) Regarding slow execution, that depends on your rasa config choices and the number of stories/rules. Finally, you need to pass the param use_gpu = True in your config for TedPolicy t make it train TED faster.