Rasa Pro new install - Illegal Instruction (core dumped) on first training

Rasa Pro CALM
1

Hello all -

Total RASA nube here, doing fresh install rasa pro on a fresh Ubuntu 22.04 image, running Python 3.10. Intention is to get environment setup to support evaluation for bigger Rasa CALM project – but can’t get started.

Installed version → (venv-rasa) jlciii@Ubuntu-2204-Rasa-Quickstart:~$ rasa --version Rasa Pro Version : 3.11.0 Minimum Compatible Version: 3.11.0rc1 Rasa SDK Version : 3.11.0 Python Version : 3.10.12 Operating System : Linux-6.8.0-49-generic-x86_64-with-glibc2.35 Python Path : /home/jlciii/venv-rasa/bin/python3 License Expires : 2027-12-11T17:53:49+00:00

When I run the initial training “rasa init”, when it goes off to run the initial training, it fails with a “Illegal Instruction (core dumped)” message. After searching around on other parts of the forum, I see a bunch of very old messages of other people encountering the same issue, and they point towards tensorflow as being the culprit.

They say to run “python3 -m tensorflow”, and sure enough, I get the same Illegal Instruction error. Same thing if I go into Python and simply run “import tensorflow as tf” - that fails too, the same way.

Other messages seem to propose that I need to downgrade my tensorflow to 1.15.0

(venv-rasa) jlciii@Ubuntu-2204-Rasa-Quickstart:~$ pip show tensorflow Name: tensorflow Version: 2.14.1 Summary: TensorFlow is an open source machine learning framework for everyone. Home-page: https://www.tensorflow.org/ Author: Google Inc. Author-email: packages@tensorflow.org License: Apache 2.0 Location: /home/jlciii/venv-rasa/lib/python3.10/site-packages Requires: absl-py, astunparse, flatbuffers, gast, google-pasta, grpcio, h5py, keras, libclang, ml-dtypes, numpy, opt-einsum, packaging, protobuf, setuptools, six, tensorboard, tensorflow-estimator, tensorflow-io-gcs-filesystem, termcolor, typing-extensions, wrapt Required-by: rasa-pro, tensorflow-text

Comment recommends downgrading tensorflow to 1.15.0, but is then corrected by a later post as something that won’t work. I’m using a fairly modern (albiet small) CPU, an Intel Core i3, with AVX support, so tensorflow likely isn’t bombing on the AVX support.

I can see a half dozen posts of this problem back in 2020-21, but none of them seem to get resolved. Is there a solution, or do I need to find new hardware to run this on?

Thanks in Advance for any help!

2

Ultimately, I was able to figure out this one by myself - but it was not straight forward to get to the root cause. As others had identified, tensor flow was the issue that was causing the problem, and it was that it specifically required AVX to be running on my system. I do have a modern Intel based laptop that does support AVX, but the issue is that I’m running inside a hosted VirtualBox environment, running my Rasa implementation as a Ubuntu image – and VirtualBox has an issue where it suppresses the AVX capabilities on systems where the host is running Windows, using Hypervisor services. Turning off Hypervisor in Windows 11 ultimately ended up being a fruitless effort, so I gave up trying to get VirtualBox to work properly.

Instead, focused on finding a version of TensorFlow that didn’t require AVX – and was able to find several references to a special build of TF that could run on these older machines. python - Run TensorFlow 2.0 on CPU without AVX - Stack Overflow

Once I was able to install a complete working stack, but replace out just the Tensorflow bits with these modified build, it worked. I suspect that it may not perform as well as the AVX version, but for the purposes of my evaluation it is working find.