How does anaconda pick cudatoolkit

Question

I have multiple enviroments of anaconda with different cuda toolkits installed on them.

env1 has cudatoolkit 10.0.130

env2 has cudatoolkit 10.1.168

env3 has cudatoolkit 10.2.89

I found these by running conda list on each environment.

When i do nvidia-smi i get the following output no matter which environment i am in

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 435.21       Driver Version: 435.21       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce RTX 208...  Off  | 00000000:01:00.0  On |                  N/A |
|  0%   42C    P8     7W / 260W |    640MiB / 11016MiB |      2%      Default |
+-------------------------------+----------------------+----------------------+

Is the cuda version shown above is same as cuda toolkit version? If so why is it same in all the enviroments?

In env3 which has cudatoolkit version 10.2.89, i tried installing cupy library using the command pip install cupy-cuda102 . I get the following error when i try to do it.

ERROR: Could not find a version that satisfies the requirement cupy-cuda102 (from versions: none)
ERROR: No matching distribution found for cupy-cuda102

I was able to install using pip install cupy-cuda101 which is for cuda 10.1. Why is it not able to find cudatoolkit 10.2?

The reason i am asking this question is because, i am getting an error cupy.cuda.cublas.CUBLASError: CUBLAS_STATUS_NOT_INITIALIZED when i am running a deep learning model. I am just wondering if cudatoolkit version has something to do with this error.Even if this error is not related to cudatoolkit version i want to know how anaconda uses cudatoolkit.

Answer 1

This isn't really answering the original question, but the follow up ones:

tensorflow and pytorch can be installed directly through anaconda without explicitly downloading the cudatoolkit from nvidia. It only needs gpu driver installed. In this case nvcc is not installed still it works fine. how does it work in this case?

In general, GPU packages on Anaconda/Conda-Forge are built using Anaconda's new CUDA compiler toolchain. It is made in such a way that nvcc and friends are split from the rest of runtime libraries (cuFFT, cuSPARSE, etc) in CUDA Toolkit. The latter is packed in the cudatoolkit package, is made as a run-dependency (in conda's terminology), and is installed when you install GPU packages like PyTorch.

Then, GPU packages are compiled, linked to cudatoolkit, and packaged, which is the reason you only need the CUDA driver to be installed and nothing else. The system's CUDA Toolkit, if there's any, is by default ignored due to this linkage, unless the package (such as Numba) has its own way to look up CUDA libraries in runtime.

It's worth mentioning that the installed cudatoolkit does not always match your driver. In that event, you can explicitly constrain its version (say 10.0):

conda install some_gpu_package cudatoolkit=10.0

what happens when the environment in which tensorflow is installed is activated? Does conda create environment variables for accessing cuda libraries just when the environment is activated?

Conda always sets up some env vars when an env is activated. I am not fully sure about tensorflow, but most likely when it's built, it's linked to CUDA runtime libraries (cudatoolkit in other words). So, when launching tensorflow or other GPU apps, they will use the cudatoolkit installed in the same conda env.