2

I am having some difficulty running code with the cudf and dask_cudf modules in python. I am working on Jupyter Labs through Anaconda. I have been able to correctly install my nvidia-gpu driver, cudf (through rapidsai), and cuda. Only, when I go to import cudf in python using import cudf, I get an error reading: "home/lib/python3.7/site-packages/cudf/utils/gpu_utils.py:120: UserWarning: No NVIDIA GPU detected. Warnings.warn("No NVIDIA GPU detected")

My environment:

  • Linux: RHEL8
  • Python: 3.7.7
  • Cuda: 10.2
  • Nvidia Driver: 390.138
  • CUDF/Dask_CUDF: 0.13 through rapidsai

I am trying to load and manipulate datasets with data in the hundreds of thousands to millions of items, so I really need the cudf/dask_cudf utility to maximize my time.

When I run nvidia-smi in the terminal, everything looks fine and the persistence mode is on. I have searched all over the internet for a solution with no great ideas. Any help would be appreciated.

rpanai
  • 12,515
  • 2
  • 42
  • 64
Maggie
  • 23
  • 5
  • 2
    Driver 390.138 is not compatible with CUDA 10.2. CUDA 10.2 requires a newer driver. I'm guessing you have an older (Fermi) GPU that you are trying to run on, you don't mention which GPU you are using or provide the `nvidia-smi` output. Rapids including CUDF [require a pascal or newer GPU](https://rapids.ai/start.html). <-- click and study the prerequisites there. – Robert Crovella Jul 13 '20 at 17:50
  • So I have tried multiple steps since this post. 1) I have tried downgrading to CUDA 9.0. 2) I have tried upgrading to the newest NVIDIA driver -450.87. (my GPU is Quadro K1200 on Linux RHEL8). The second option causes my system to crash and I have to ssh into the computer to remove nvidia and get it working. The first option results in the same cudf/dask_cudf error. Trying other alternatives to CUDA results in a driver/library mismatch error. – Maggie Jul 13 '20 at 21:47
  • So there's no way to use CUDF unless I actually get a new GPU? – Maggie Jul 13 '20 at 22:02
  • Sorry, as it happens K1200 is a maxwell GPU, but that also predates Pascal, the stated minimum GPU for RAPIDS. It's an open-source project. If you are interested in debugging problems that nobody has tested before, and making your own changes to the CUDF codebase, and have some expertise with CUDA, it may be possible to use CUDF with your GPU. However, I'll partially repeat myself: I don't think your setup is a good starting point for Rapids/CUDF usage. It doesn't meet the stated minimum. But if you are sufficiently motivated, good luck! Give it a try. – Robert Crovella Jul 13 '20 at 22:13
  • FWIW if you are attempting to use any kind of binary install (e.g. Anaconda, or pip) for any recent version of RAPIDS, I don't think it will work. You'll need to recompile from sources. – Robert Crovella Jul 13 '20 at 22:14

1 Answers1

2

Based on the conversations you're having with Robert, it seems that your GPU's architecture being a few generations outside of what RAPIDS will work with is the issue. Thanks Robert for working with Maggie to figure that out!

I wouldn't try to force RAPIDS to work on Kepler when there are so many alternative ways to provision a GPU - even free options for trial purposes!

If you are still interested in trying out RAPIDS and only need a single GPU, please look at our Google Colab notebooks and set up script OR app.blazingsql.com. They are shared or extra instances, with Colab allowing you more customization of your workspace if you need to install more packages and blazing having the fastest "get up and running" time.

If you feel that you need more than one GPU, you move to the paid realm and can provision it with any major cloud provider, install RHEL version of your choice (we only officially support RHEL 7, though).

Does that help you?

TaureanDyerNV
  • 1,208
  • 8
  • 9