6c64741 introduced an error moving an fp16 pipeline to CPU, which makes sense as it cannot run on CPU; however I would suggest changing it to a warning as it broke my use-case that I think is a legitimate one
This is the use-case that broke:
- Load a bunch of different pipelines running different models on CPU (RAM), but never actually run them for inference on CPU
- Whenever I want to run them for inference, I move them from GPU -> CPU which is an operation that is way faster than if I had to load them from disco every
- Once the model is ran, the pipeline with the model loaded is sent back to CPU (RAM) to be idle but warm