Rasa with CUDA 10.2

Rasa Open Source
1

Hi,

I am currently working on a university server that has CUDA 10.2 installed and I have no means to install other CUDA versions (I assume there is no way without root rights?).

When using rasa, I get the following warning, so I cannot use GPU to train the rasa model:

(DL) [rasa]$ python
Python 3.6.12 |Anaconda, Inc.| (default, Sep  8 2020, 23:10:56) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import rasa
2021-01-05 12:22:48.411064: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'libcudart.so.10.1'; dlerror: libcudart.so.10.1: cannot open shared object file: No such file or directory
2021-01-05 12:22:48.411122: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

Aparently this is due to an incompatibility of TF 2.3 and CUDA 10.2. Unfortunately, rasa requires TF<2.4.

Would you know a workaround for this? When is TF 2.4 going to be supported?

1 Like
2

We’re aware of the issue and there’s a github issue for this here. I can’t make any promises on when it will be added though.

That said, my colleague @dakshvar22 suggests that CUDA 10.2 should be usable with tf==2.3. It’s only cuda 11 onwards which tf 2.4 supports.

3

Hi @ValeKnappich,

is it important to actually use a GPU or would it be OK to use CPU?

If you are fine to use CPU for the moment, you could set the environment variable:

CUDA_VISIBLE_DEVICES=""

Kind regards
Julian

4

@koaning @JulianGerhard

Thanks for your quick answers.

Aparently, tensorflow skipped support for 10.2, since they directly jumped from 10.1 to 11. Anyway, for future readers, the issue can be solved by symlinking the 10.2 libraries to 10.1, since their apis are compatible. (See this issue)