Efficient KeyLLM + KeyBERT - Torch not compiled with CUDA enabled

Testing out Efficient KeyLLM + KeyBERT in a pipeline, when the code hits the refinement stage using LLM I get error

"Error while extracting keywords from text: Torch not compiled with CUDA enabled"

This line that breaks the code inside extract keywords class - `doc_embeddings = torch.from_numpy(doc_embeddings).float().to("cpu")`

Start with the extract keywords class and reproduce the Efficient KeyLLM + KeyBERT pipeline at the refinement stage, focusing on the reported conversion to CPU. Confirm the failure and determine the relevant device-handling path; done means the pipeline no longer raises the CUDA compilation error in the reported scenario.

Written by the indexing model from the issue text.