Skip to content

Ubuntu & CUDA Setup

Hardware

GPU:    NVIDIA GeForce RTX 4050 Laptop GPU (6GB VRAM, Ada Lovelace architecture)
OS:     Ubuntu 24.04.3 LTS (Noble Numbat)
Kernel: 6.17.0-14-generic
Driver: 590.48.01

Verify your driver

If you're setting this up fresh, install the NVIDIA driver first. On Ubuntu 24.04:

sudo apt update
sudo ubuntu-drivers install
sudo reboot

After reboot, verify:

nvidia-smi

You should see something like this:

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.48.01    Driver Version: 590.48.01    CUDA Version: 13.1               |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                Persistence-M  | Bus-Id          Disp.A | Volatile Uncorr. ECC |
|   0  NVIDIA GeForce RTX 4050 ...   Off  | 00000000:01:00.0   On  |                  N/A |
| N/A   38C    P8    2W / 60W             |    60MiB / 6141MiB     |     39%      Default |
+-----------------------------------------------------------------------------------------+

Three things to check:

  • Driver version is listed — means NVIDIA driver is loaded correctly
  • GPU name appears — means the GPU is detected
  • Memory shows available VRAM — means the GPU is healthy

The CUDA situation — read this carefully

This tripped us up and is worth understanding properly.

Running nvidia-smi shows CUDA Version: 13.1 in the top right. This is not the CUDA toolkit version — it's the maximum CUDA version your driver supports. It's a capability ceiling, not what's actually installed.

When we ran:

ls /usr/local/ | grep cuda
# cuda
# cuda-13
# cuda-13.0

We found CUDA 13.0 toolkit installed. That's very new. PyTorch stable builds don't have wheels for CUDA 13.x yet.

But here's the key insight: PyTorch doesn't use your system CUDA toolkit.

PyTorch ships with its own CUDA runtime bundled inside the wheel. When you install PyTorch, it brings libcudart, libcublas, and everything else it needs. Your system CUDA toolkit is only relevant if you're compiling custom C++/CUDA extensions from source.

So the actual CUDA situation for this project is:

What Version Where
System CUDA toolkit 13.0 /usr/local/cuda-13.0
PyTorch bundled CUDA runtime 12.8 Inside PyTorch wheel
CUDA your code actually runs on 12.8 PyTorch's bundled runtime

Your driver (590.x) supports CUDA runtime 12.8 without any issues — driver support is forward-compatible within major versions.

Verify CUDA is working with PyTorch

After installing PyTorch (covered in the next section), run:

import torch

print(torch.__version__)         # 2.10.0+cu128
print(torch.cuda.is_available()) # True
print(torch.cuda.get_device_name(0)) # NVIDIA GeForce RTX 4050 Laptop GPU
print(torch.cuda.get_device_properties(0).total_memory // 1024**3) # 6 (GB)

All four lines should return meaningful values. If torch.cuda.is_available() returns False, the driver is either not installed or not visible to PyTorch.

cuDNN

cuDNN 9.13.1 came pre-installed on this system alongside CUDA 13.0:

dpkg -l | grep cudnn
# cudnn9-jit-cuda-13  9.13.1.26-1

PyTorch also bundles its own cuDNN, so this is again not directly used by PyTorch — but it's good to have for future custom kernel work.

What's next

With the driver verified and CUDA confirmed working through PyTorch, we set up Python and the package manager. → Python & uv