I am doing a fresh installation of rasa 2.2.0 on Ubuntu 20.04 with Lambda Stack installed. This includes CUDA v11.1, cuDNN v7.6.5, and TF v2.3.0. Any rasa command throws a dlerror for libcudart.so.10.1. It looks like rasa depends on CUDA 10. Installing from source doesn’t change this.
Further evidence: a virtual environment before pip3 install rasa successfully accesses libcudart.so.11.1; after the command, the above mentioned error is thrown.
Does rasa have a hard dependency on CUDA 10? Can it run TF on CUDA 11 instead? If yes, how should it be configured?
@ivogeorg As mentioned here CUDA 11 is only supported for TF>=2.4.0. Since rasa==2.2.0 requires tensorflow~=2.3.0, I would recommend you to downgrade CUDA to 10.1 or similar.
@Ghostvv is it possible to build rasa with a specific cuda toolkit version? I am very new to poetry build tool, but I see that rasa/poetry.lock at main · RasaHQ/rasa · GitHub mentions cuda* extras from rasa’s dependencies spacy and thinc. And, for example, I can install spacy with a support to a specific cuda version using pip: pip install -U spacy[cuda92]. It would be great to be able to build rasa with an extra build parameters like cuda92. Thanks!