NVIDIA RTX 5060 8GB smi 120 problem fixed (is this is helpfull good)

[victorjavierdavalos123-coder]

Problem

If you have an RTX 5060/5070/5080/5090 (Blackwell architecture), the project will fail to run because:

1. The default installation uses torch with CUDA 12.6 which does NOT support sm_120 (Blackwell)
2. When manually installing packages with pip install --target, pip uses your system Python (3.11) instead of the project's portable Python (3.10), causing all C-extension libraries (numpy, scipy, pillow, torch) to be compiled for the wrong Python version

Symptoms

• ModuleNotFoundError: No module named 'torch'
• ImportError: No module named 'numpy.core._multiarray_umath'
• ImportError: cannot import name '_imaging' from 'PIL'
• ImportError: scipy install seems to be broken

Solution

Instead of using plain pip install, always use the project's portable Python executable directly. This ensures all packages are installed for the correct Python version (3.10) and in the correct location.

Step 1 - Install torch with CUDA 12.8 support (required for Blackwell/sm_120):

.\system\python\python.exe -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 --upgrade

Step 2 - Reinstall numpy for Python 3.10:

Download the correct wheel manually from https://pypi.org/project/numpy/1.26.4/#files
Look for: numpy-1.26.4-cp310-cp310-win_amd64.whl

Then install it:

.\system\python\python.exe -m pip install "PATH\TO\numpy-1.26.4-cp310-cp310-win_amd64.whl"

Step 3 - Reinstall pillow and scipy:

.\system\python\python.exe -m pip install pillow scipy --upgrade

Step 4 - Install remaining torch dependencies:

.\system\python\python.exe -m pip install torchsde torchmetrics torchdiffeq --upgrade

Step 5 - Verify everything works:

.\system\python\python.exe -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

Expected output: 2.7.0+cu128
True
NVIDIA GeForce RTX 5060 Then run the project normally:

.\run.bat

Key rule

Always use .\system\python\python.exe -m pip install instead of plain pip install when working with this project. Using plain pip will install packages for your system Python instead of the project's portable Python 3.10.

Tested on

• GPU: RTX 5060
• CUDA: 12.8
• Torch: 2.7.0+cu128
• OS: Windows 11