Runnying Jupyter server on Windows

2026-01-08

I have a Windows machine with an RTX 3090, which I’d like to use for some local AI training. I still want to run the notebook on my MacBook, however. Here’s what I had to do to make that happen.

CUDA

For stability, install Nvidia Studio driver instead of Game Ready. It’s the same tech stack, but Studio is more stable. Restart the computer afterwards.

Made this change after seeing a Jupyter kernel crash & restart:

[I 2026-01-09 09:37:14.194 ServerApp] Uploading file to /252_28238_29-pycharm-support-libs.zip
[I 2026-01-09 13:07:44.389 ServerApp] AsyncIOLoopKernelRestarter: restarting kernel (1/5), keep random ports
[W 2026-01-09 13:07:44.390 ServerApp] kernel 2fb85dd2-84c3-49dc-9395-26c8dfdbee4b restarted

Checked in Windos Event Viewer for an error:

Faulting application name: python.exe, version: 3.11.8150.1013, time stamp: 0x65c2ad47
Faulting module name: c10_cuda.dll, version: 0.0.0.0, time stamp: 0x67185a9f
Exception code: 0xc0000005
Fault offset: 0x0000000000021a5d
Faulting process id: 0x3dd8
Faulting application start time: 0x01dc818e95dab2f7
Faulting application path: C:\Python311\python.exe
Faulting module path: C:\venvs\jupy\Lib\site-packages\torch\lib\c10_cuda.dll
Report Id: f10c260f-be0b-4d0b-b829-f0e3fda6b08a
Faulting package full name:
Faulting package-relative application ID:

Python

First, check CUDA version with nvidia-smi:

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 566.03                 Driver Version: 566.03         CUDA Version: 12.7     |
|-----------------------------------------+------------------------+----------------------+

Need to install a CUDA-enabled version of torch from the right index, https://download.pytorch.org/whl/cu1211. No compatible torch wheel existed for Python 3.13, which we already had. So installed 3.11:

choco install python --version=3.11.8 -y

Then created a virtual environment called jupy for installing dependencies:

py -3.11 -m venv C:\venvs\jupy

Activate the venv:

C:\venvs\jupy\Scripts\Activate.ps1

Then install ipykernel:

pip install ipykernel

and register the kernel:

python -m ipykernel --user --name jupy --display-name "Python (jupy)"

Jupyter

Clone the GitHub repo you want to work in. It’s easier if the server is started in that directory.

Use a password to access the server. Run:

jupyter notebook password

Set the HuggingFace token as environment variable HF_TOKEN. Start a new PowerShell window. Start the server:

jupyter notebook --ip=0.0.0.0 --no-browser

Find the Windows machine’s IP address:

ipconfig

Wireless LAN adapter Wi-Fi:
   ...
   IPv4 Address. . . . . . . . . . . : 192.168.10.104
   ...

On the Mac, go to System Settings > Privacy & Security > Local Network and enable Chrome / Pycharm to communicate with devices on local network. Then go to http://192.168.10.104:8888 and see that the server is accessible.

PyCharm

In PyCharm, go to Jupyter Servers and add a new server. Choose Notebook/Lab, provide URl and password. Test connection to verify it works. Connect to server and switch to jupy kernel.

When installing dependencies, use {sys.executable} to use the Python executable in the virtual env:

import sys
!{sys.executable} -m pip install torch --index-url https://download.pytorch.org/whl/cu121
!{sys.executable} -m pip install transformers datasets peft accelerate

Footnotes

  1. cu121 stands for CUDA 12.1, which is compatible w/ 12.7. ↩