GPU and CUDA

GPU and CUDA

Whisper (faster-whisper / CTranslate2) and diarization (pyannote / PyTorch) share one device per job: both GPU or both CPU — never mixed.

Install CUDA stack (recommended)

lwt install cuda

This installs:

1. CUDA 12 runtime (nvidia-cublas-cu12, nvidia-cudnn-cu12) — Whisper GPU
2. CUDA PyTorch (torch+cu126) — diarization neural nets on GPU
3. CPU torchaudio — resample/fbank helpers (CUDA torchaudio is often blocked on Windows)

Optional full toolkit (Windows):

lwt install cuda --toolkit

Check

lwt install check

Look for CUDA ready for Whisper + diarization.

Behaviour

Mode	Device
Transcribe only	GPU if CTranslate2 sees CUDA
Transcribe + --diarize	GPU only if both CTranslate2 and PyTorch see CUDA; otherwise CPU for both

Force CPU:

lwt config set whisper.device cpu