Unable to find GPU on Windows

[sarthakpati]

Hi,

I'd like to thank and commend you on putting this together!

I am running Windows and this is my output of nvidia-smi:

(base) PS C:\Users\sarth> nvidia-smi.exe
Tue Jul 28 16:16:35 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 451.77       Driver Version: 451.77       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce RTX 208... WDDM  | 00000000:01:00.0  On |                  N/A |
| N/A   46C    P8     7W /  N/A |   4402MiB /  8192MiB |     18%      Default |
+-------------------------------+----------------------+----------------------+

But, I am not able to detect the GPU from the GPUtil:

>>> os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
>>> GPUtil.getAvailable()
[]
>>> GPUtil.__version__
'1.4.0'

Is there something extra I need to add in the python code to get this working?

Thanks!

[jamieslyman]

On my machine, the executable existed in System32, but the NVSMI folder did not exist in Program Files\NVIDIA Corporation and spawn couldn't find the executable, despite it being in System32.
I fixed it by copying the NVSMI folder from an old install that had it.
If you can find them in your System32, the files are MCU.exe, nvdebugdump.exe, nvidia-smi.1.pdf, nvidia-smi.exe, and nvml.dll
I don't know what files are required but I'd guess the absolute minimum would be nvidia-smi.exe and nvml.dll.

[gchennell]

Copied just the nvidia-smi.exe and nvml.dll and it seems to have resolved the lack of stats - thanks @wdcook4

[sarthakpati]

Copied just the nvidia-smi.exe and nvml.dll and it seems to have resolved the lack of stats - thanks @wdcook4

Copied them to where?

[gchennell]

I had an empty nvsmi folder which I copied these in to. I think if you don't have the nvsmi folder just create it and see...

[sarthakpati]

Is it at a particular location relative to the module or should it be in the PATH?

[gchennell]

Hi, the same as the earlier comments. c:/program files/NVIDIA Corporation/NVSMI yes?

…

On Mon, 8 Feb 2021 at 13:07, Sarthak Pati ***@***.***> wrote: Is it at a particular location relative to the module or should it be in the PATH? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#37 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB46Y5QE72GEYKKJ673RDP3S57OXHANCNFSM4PK4JVKA> .

-- Sent from my PC

[sarthakpati]

Right, thank you. I already have that present, so unsure what's going on...

[Apfelkuchenbemme]

Little late to the party but this probably has nothing to do with nividia-smi not being found, though just setting it to nvidia_smi = "%s\\Program Files\\NVIDIA Corporation\\NVSMI\\nvidia-smi.exe" % os.environ['systemdrive'] without a check if that file exists is pretty wild, not to mention handling the exception at the Popen call by silently returning an empty list.

Instead, you have to take a very, verrryy close look at your output for (base) PS C:\Users\sarth> nvidia-smi.exe, where you'll find that you used 4402MiB out of 8192MiB, which means at the time of testing, you used 4402/8192 ~ 53.7% of available memory.

The getAvailable()-function itself first calls getGPUs() with the aforementioned "interesting" solution to spawn not finding nvidia-smi. Then, in line 143, it calls getAvailability(..) to which we pass the default arguments you called getAvailable() with.

One of these arguments is maxMemory=0.5. You can probably see where this is going already, but with your aforementioned 4402MiB of used memory out of 8192MiB total memory, you're just above the 50% memory used threshold, which means line 177 puts a 0 for the "availability" of your RTX2080, thus getAvailable() removes it from the list, ultimately returning an empty list.

[rayanelahmadi]

Right, thank you. I already have that present, so unsure what's going on...

Did you ever figure out the issue? I have the same problem.

[sarthakpati]

Unfortunately, not. Since I am using PyTorch for most of my work, I am no longer reliant on this specific library (docs for PT Cuda are here).

If you do manage to figure out the issue, please do LMK so that I can use it for some of my non-PT work. 👍🏽

[rayanelahmadi]

Unfortunately, not. Since I am using PyTorch for most of my work, I am no longer reliant on this specific library (docs for PT Cuda are here).

If you do manage to figure out the issue, please do LMK so that I can use it for some of my non-PT work. 👍🏽

So, I actually figured out my issue. I had an old, 32-bit version of Python installed and so I went ahead and uninstalled it, installing a newer 64-bit version (Python 3.12 64-bit). I then re-installed GPUtil and tried everything again (I think it also made me pip install setuptools). Everything now works.