Expected CPU usage when using GPU

[filmo]

I've done all the things to set pipeline.to(torch.device('cuda:n')) where n is my GPU number.

It appears to load the model onto the GPU (I see a process with ~1 to 2 GB occupancy ) and it does perform the diarization. About 30 seconds for a 30 minute file.

However, I also see that during the pyannote process, it's essentially maxing out my CPU (i5 13600k) while performing the diarization.

Is this expected behavior when the Pyannote is "running on the GPU" ??

I'm running a ASR chain that involves

• NEMO MSDD diarization
• pyannote diarization.
• faster-whisper
• whisperX
• a punctuation model

The problem is if I run two separate ASR chains (each on a separate RTX-3090), the pyannote process slows down dramatically as the available CPU compute is split between the two processes. Performance drops from ~30 seconds down up to ~280 to 320 seconds. (all else being equal, I would have thought splitting the CPU compute would double or triple the time (60 to 90 seconds) but not by an order of magnitude.

Such high CPU usage makes me suspect that my pyannote may actually be running on the CPU even though it appears the model is loaded onto the GPU. I'm now very confident that it is on the GPU as I forced it to CPU and the runtime was fantastically longer.

I'm still curious about the high CPU usage even when running the model on the GPU.