PyTorch CUDA Availability Issue on NVIDIA GeForce MX450

1

Hi there.
Despite having a CUDA-enabled version of PyTorch installed, torch.cuda.is_available() is returning False.
Could I please seek your assistance with this?
Thanks in advance.

import torch
print("PyTorch version:", torch.__version__)
print("CUDA version:", torch.version.cuda)
print("Is CUDA available:", torch.cuda.is_available())

When I execute the Python code above in JupyterLab notebook, I get the following output:

  • PyTorch version: 2.2.2+cu121
  • CUDA version: 12.1
  • Is CUDA available: False

According to the Device Manager on my personal laptop, I have the following specifications:

  • GPU: NVIDIA GeForce MX450
  • Driver date: August 1, 2021
  • Driver version: 27.21.14.5774

It also states that “The best drivers for your device are already installed”.

2

This does not seem to be an NVIDIA driver based on the version, so you might need to reinstall it.

3

Thanks @ptrblck for your reply.
The left snapshot displays the system information from the NVIDIA Control Panel on my laptop. I’ve tried to download several NVIDIA drivers from https://www.nvidia.com/Download/Find.aspx?lang=en-us# (middle snapshot), but none of them have passed the compatibility check so far (right snapshot). Additionally, torch.cuda.get_arch_list() returns an empty list [].
I’m unsure what I might have done incorrectly.

Screenshot 2024-04-03 154242
Screenshot 2024-04-03 1542421920×716 109 KB

4

Once you’ve properly installed a driver, double check that a single PyTorch binary with CUDA support is installed since the current one does not seem to ship with any CUDA kernels and thus returns an empty architecture list.

5

Hi @ptrblck. Thanks again for your response.

I uninstalled the current version of PyTorch and reinstalled it using the following command:

!pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

I followed the instructions provided at Start Locally | PyTorch.

The following Python code returns True:

import torch
torch.cuda.is_available()

Additionally, when I execute the following command:

torch.cuda.get_arch_list()

I receive the following output:

['sm_37', 'sm_50', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90', 'compute_37']

print(torch.cuda.get_device_name(0))  # Returns 'GeForce MX450'

I believe everything is working correctly now.

6

Yes, now it looks good. Run a quick test by allocating a random tensor on the GPU via torch.randn(1).cuda() and make sure you are able to allocate this tensor.

7

Hi @ptrblck .
When I run torch.randn(1).cuda(), it returns tensor([0.1023], device='cuda:0').
Thanks for the advice!