When deploying an interface to call llamafactory on a Huawei NPU server, a problem arose where the specified graphics card could not be called correctly.

[DavidCam3]

I deployed the API in a Docker container, specifying GPUs 6 and 7 for fine-tuning training, but these two GPUs could not be used correctly. Training worked fine on GPUs 0 and 1, but GPUs 0-3 had processes running, indicating they were unusable. Calling GPUs 6 and 7 resulted in an error message saying "Device not found." I've tried almost everything I could think of. What could be causing this?