According to this https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html my installation seems to meet minimum requirements.
However, I receive “CUDA failed with error CUDA driver version is insufficient for CUDA runtime version”
Please suggest the solution. I’m out of ideas.
Additional info
There is some aspect of your development toolchain that requires something higher than CUDA 11.4 support. It might be something other than the CUDA 11.4 installation that you show the nvcc output of. For example, another CUDA installation, a conda environment, or some other installed toolchain like tensorflow or pytorch.
A usual suggestion in this case that I offer is simply to upgrade your GPU driver to the latest one available for your GPU.
The alternative (perhaps) is to identify the application that generated the runtime message:
and carefully determine the exact tools that were used to build it. Even then, it may be necessary to update the GPU driver to resolve.
This is top driver version already.
Do you think compatibility upgrade could help?
Correct, so you are attempting to use some kind of toolchain which simply cannot be used with Tesla K80. That GPU is too old.
No, I don’t expect compatibility libraries would help. You’re welcome to try whatever you wish, of course. Perhaps there is an older version of the code that you are trying to use, that could be run with an older toolchain. Or you may wish to get access to a newer GPU.
But in general … Please confirm such installation should work …
Ubuntu 20.04
Tesla K80
Driver 470.256.02
Cuda 11.4
pytorch 2.0.1
I will reinstall it then.
This is not a forum for help with pytorch. I generally wouldn’t expect any recent version of pytorch to work with a Tesla K80. I won’t be able to help further there.
If you want to validate that your CUDA install is working correctly, instructions are provided in the linux install guide. Beyond that, if your CUDA install is working correctly, but pytorch isn’t, then you will need to look at things from the pytorch angle.
It looks to me like pytorch 2.x.x requires CUDA 11.7 as a minimum install prerequisite. So I wouldn’t expect pytorch 2.0.1 to be “easily usable” with CUDA 11.4. Beyond that, I very much doubt that pytorch 2.x installer packages from pytorch are natively built with support for cc3.7 GPUs like the K80.
And the error I received is not about pytorch.
It says:
CUDA failed with error CUDA driver version is insufficient for CUDA runtime version
Thanks anyway.
That error report is coming from code that is part of pytorch.
If you stop using pytorch, and otherwise start using CUDA correctly on that platform (e.g. with codes that you compile with the nvcc version you posted in your original message) then I expect you will not see that error message, and your codes should work normally.
The problem is with the specific version of pytorch that you are using, and its incompatibility with the specific version of CUDA that you are using. And as we have seen, it cannot be fixed from the CUDA side.
OK
Thanks)
