[BUG] LLMLingua fails on Windows with PyTorch CPU: "Torch not compiled with CUDA enabled"

Description:
I'm trying to use LLMLingua on Windows 10 with PyTorch CPU (no GPU/CUDA available).
When initializing the model microsoft/llmlingua-2-xlm-roberta-large-meetingbank (or any other), I get the following error:

AssertionError: Torch not compiled with CUDA enabled

Steps to reproduce:

1. Install PyTorch CPU (pip install torch --index-url https://download.pytorch.org/whl/cpu)
2. Install llmlingua and transformers (pip install llmlingua transformers)
3. Run:

   from llmlingua import PromptCompressor
   compressor = PromptCompressor(model_name="microsoft/llmlingua-2-xlm-roberta-large-meetingbank", use_llmlingua2=True)
   

4. The error occurs even if os.environ['CUDA_VISIBLE_DEVICES'] = '-1' is set before any import.

Environment:

• Windows 10
• Python 3.11
• torch 2.6.0+cpu
• llmlingua 0.2.2
• transformers 4.51.3

Full error output:

AssertionError: Torch not compiled with CUDA enabled

(The traceback points to an internal call in transformers/llmlingua trying to access CUDA.)

Notes:

• The environment has no GPU or CUDA.
• The error occurs even with the latest version of all libraries.
• The same code works on Linux (according to other users).

Question:
How can I force LLMLingua to use CPU only on Windows?
Is there any workaround or model version that works on pure CPU in Windows?

Thanks for your help!