Nvidia GPU and CUDA Notes

SD.Next auto-detects nVidia GPUs and attempts to install a stable CUDA-enabled torch build at launch.

Important

Do not manually install CUDA on your system as torch will install the appropriate CUDA version in its own environment
System-wide installation of CUDA may cause conflicts and is not required for SD.Next to utilize nVidia GPUs.

[!TIP] Disable system memory fallback to avoid massive performance degradation when exhausting GPU memory: Instructions

Older GPUs

If you have an older nVidia GPU that is not supported by recent torch releases, you may need an older torch and CUDA combination.

Example: Pascal architecture (for example, RTX 10xx series) or older may require torch==2.9 with cuda==12.6.
Use the following before first startup to force a specific torch/CUDA combination:

set TORCH_COMMAND='torch==2.9.1 torchvision==0.24.1 torchaudio==2.9.1 --index-url https://download.pytorch.org/whl/cu126'

Older Drivers

If your system has older nVidia drivers that do not support newer CUDA versions, you may need an older torch/CUDA pair compatible with those drivers.

This is common when running on cloud instances which do not update drivers frequently.

Example: cuda==13.0 requires newer nVidia drivers.
Use the following before first startup to force installation from CUDA 12.8 wheels:

set TORCH_COMMAND='torch torchvision --index-url https://download.pytorch.org/whl/cu128'