**Describe the bug** Input of promt_encoder is on wrong device if more than 1 cuda device is available. Inference Fails. **To Reproduce** Run fully integrated pipeline using the slicer plugin on machine with more than 1 gpu