Some Notes on & Reached Dead End in WSL2 Install

Hello all,

I am trying to install WSL2 on my Windows 11/Ubuntu 20.04.5 machine using the instructions linked to in this TF Forum post. It appears there are some known issues with installing this release that are discussed at this GitHub post, which some might find helpful to know about.

Per the first recommendation of the post, I downgraded the TF version I was installing to 2.9.2 to solve the issues mentioned. At the bottom of the post he states that it is also possible to solve the issue by upgrading to cuDnn 8.2.4. Since I still received one more error with the install of the downgraded version of TF flow, I am wondering if all problems would be solved by installing the upgraded version of 8.2.4. If so, would it be possible to install it using conda install c-xxx cudnn=version syntax by adding a new channel? Or if I have to go somewhere else and install it, what would the commands be for doing so and where would one go?

So sticking with the 2.9.2 version still has a problem. The error is: “tensorflow/stream … could not open file to read NUMA mode: … Your kernel may have been built without NUMA support.” I went to check if the WSL kernel was updated and had NUMA support. I updated it, but per this command: numactl --hardware, I was told “No NUMA available on this system”.

So now I am at a loss as to what to do next… how can I install the updated version of Cudnn for WSL and what can you do when the kernel doesn’t seem to support NUMA and that is the exact error thrown? Any help as to what the next steps may be is greatly appreciated. Thanks!

Hello all,

Here is an update:

The NUMA support comment is a warning, not an error, so TensorFlow still builds. Thus, I am able to build on TensorFlow 2.9.2 and run the GPU version. Any insight on the warning and why it pops up is welcomed.

I am still unable to solve the problem mentioned in the post with TensorFlow 2.10 and am still interested in trying to download and install the higher version of cudnn, version 8.2.4, to see if I can start working with TensorFlow 2.10. So I am still requesting direction here. Note it would have to be for WSL.