Error messages

Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

You run webui-user.bat (or webui.sh), the launcher installs or checks its packages, and then it quits with a traceback ending in:

RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

The WebUI's own troubleshooting wiki calls it "one of the most frequently mentioned problems" and says it's "usually not a WebUI fault." As of October 1, 2026 there were 148 issues containing the phrase in the AUTOMATIC1111 repo. The oldest big thread, #1742, has 270 comments.

What the launcher checks

We read modules/launch_utils.py at v1.10.1. During setup it runs one line of Python in the WebUI's own environment:

import torch; assert torch.cuda.is_available()

If that fails, you get the message. So the question is narrow. Can the copy of PyTorch inside the WebUI's venv folder (its private Python environment) find a working CUDA device? CUDA is Nvidia's GPU platform, and the default install is a CUDA build of torch 2.1.2. Any other answer stops the launch.

If you have an Nvidia card

Then something is wrong with the install, and the wiki's advice is to delete the venv folder and launch again so it's rebuilt. That's also what a collaborator told someone whose torch broke after installing the wav2lip extension (#15124): remove the extension and "Delete the venv folder itself." The wiki also says to undo whatever you changed last if it started after an update.

On a laptop or a desktop with integrated graphics, collaborator missionfloyd suggested in #16941 (April 2025) that "It may be trying to use the CPU's integrated graphics," and to try --device-id 0 or --device-id 1, or turn the integrated GPU off in the BIOS.

If it still fails, the wiki asks for the output of python -m torch.utils.collect_env, run from a console in venv\Scripts. It lists the torch version with the CUDA version it was built for, whether CUDA is available, and the Nvidia driver it found.

AMD, Intel, Mac and CPU-only

Here the message is expected. CUDA only exists on Nvidia hardware, and the wiki says to add --skip-torch-cuda-test and follow the install guide for your hardware. Adding the flag by itself won't make an AMD card work.

• AMD on Windows: collaborator w-e-w answered an RX 6800 owner in #16861 (February 2025) with "AMD + Windows not supported, only on linux," and pointed to lshqqytiger's DirectML fork. Bugs in that fork go to its own repo.
• AMD on Linux: webui.sh detects an AMD GPU and installs a ROCm build of torch (ROCm is AMD's answer to CUDA).
• Intel Arc: the --use-ipex option skips this test on its own.
• No usable GPU: the wiki's CPU recipe is --use-cpu all --precision full --no-half --skip-torch-cuda-test, and it warns that it's very slow.

Flags go on the set COMMANDLINE_ARGS= line of webui-user.bat, or the COMMANDLINE_ARGS line of webui-user.sh.

Forge prints something else

Forge, lllyasviel's fork, kept the same check but swapped the text for "Your device does not support the current version of Torch/CUDA! Consider download another version:" with a link to its releases page. The old line is still in its code as a comment, and 13 issues in the Forge repo contain it.