TF 2.20 C API (libtensorflow.so)

TensorFlow General Discussion
,
1

I spent two days trying to build a TF 2.20 library. I understand that from 2.19 no pre-compiled binaries are made, but apparently it is practically impossible to do it manually either.
I used TF 2.20, used bazel 7.4.1 and I have cuda 13.1, I have a long history of what I tried, though I do not think it is very interesting. All sort of problems with absolute and relative paths, C compiler versions, etc.
Is there a clean way to build it? Has anyone did it?

The problem is that my RTX5080 cannot use older cuda-s (I made it running with 12.5, but I have constant and unpredictable errors. The same code in two subsequent runs give different errors, like floating point error, division by zero, etc.

Thanks,

2 Likes
2

Hi @Zsolt_Szakaly, Thank you for reporting this issue.
Could you please take a look into this issue1 and issue2 and let me know if it helps.

3

Thanks for the links. However those are more related to the other end of the problem, i.e. what version can run on an RTX50x0. My problem also started somewhere there, and I realised that older versions (I tried TF2.18 with Cuda 12.5) have some issues (errors at random places, probably due to some parallelisation) and so I want to use TF2.20.
This is where the real problem starts. TF2.20 is not offered with a libtensorflow.so binary. It is said to be built from source, what I simply could not do. I tried all possible ideas Gemini had :slight_smile: , but no success. So, I am looking for either a trustworthy place to get the binary or a clear, working instruction to build it myself.

1 Like