I’m working doing infra support to 17 Lambda servers we have in a datacenter but I really don’t know how the lambda stack is actually implemented, looks like installing CUDA through the oficial NVIDIA instruction for Ubuntu 18.04 creates a folder /usr/local/cuda-10.2 with a symlink /usr/local/cuda, this is something that the developers are used to use but they can’t find it. For what I’ve seen from how the lambda stack implements cuda there doesn’t seem to be a way to get all cuda libraries and binaries in one single folder like /usr/local/cuda, am I missing something? how can I actually get this /usr/local/cuda folder with the Lambda stack without installing the official NVIDIA CUDA package?
2 Likes
Any update on this?
I just ran into the same issue. Another program I’m trying to compile is looking for libraries in /usr/local/cuda which now doesn’t exist when I installed CUDA using lambda stack. I also didn’t find any documentation on how lambda stack is implemented and where the CUDA libraries are.
It would be nice to know if uninstalling lambda stack and installing CUDA through NVIDIA is the “only” way around this. Everything else works great through lambda stack so it would be unfortunate.
1 Like
@finn I’ve personally had nothing but trouble with Lambda Stack, especially if I needed a particular version of something for any reason. Ever since I uninstalled it and installed everything as intended by the respective maintainers, I’ve had no more weird issues.
It doesn’t help that there isn’t much coming from Lambda support staff on these forums. I’ve had better support from them via email, I’d recommend trying that if you have an issue.
2 Likes
Lambda Stack is designed to support people training with PyTorch and Tensorflow using the latest version of CUDA. If you need an older version, just install that into /usr/local/cuda-9.0, for example, and link to that. Most software that you’re compiling from source will support the latest version of CUDA and work with Lambda Stack.
Finn, the header files are not stored in /usr/local/cuda/include for Lambda Stack but in /usr/include/
1 Like
I created a symbolic link:
ln -s /usr /ust/local/cuda
it worked. But I can’t find cudnn.h on my entire system?
