Troubleshooting
Solve common GPU acceleration issues.
GPU Detected But Using CPU
Symptom: Output shows Device: cpu even though you installed [gpu]
Solution:
- Verify CUDA packages are installed:
python -c "import nvidia.cublas.lib; import nvidia.cudnn.lib; print('✓ GPU packages OK')"
- Check your venv has GPU support:
pip list | grep nvidia
Should show nvidia-cublas-cu12 and nvidia-cudnn-cu12
- Verify GPU is detected by system:
nvidia-smi
Should list your GPU
- If still using CPU, reinstall GPU extra:
pip install --upgrade voicepad-core[gpu]
"CUDA 13 Incompatible" Error
Symptom: System has CUDA 13 installed, but voicepad mentions CUDA 12
Why this happens: faster-whisper requires CUDA 12, but your system has CUDA 13
Solution: No solution needed! The [gpu] extra installs CUDA 12 in your virtual environment, separate from system CUDA 13. They don't conflict.
- System keeps CUDA 13 globally
- Your venv uses CUDA 12 packages
- Both coexist without problems
GPU Not Detected
Symptom: voicepad config system shows GPU as not available
Causes & Solutions:
| Issue | Check | Solution |
|---|---|---|
| Old GPU | Compute Capability < 6.0 | Upgrade GPU (GTX 10-series minimum) |
| Old driver | nvidia-smi --query-gpu=compute_cap |
Update NVIDIA drivers |
| No CUDA packages | pip list \| grep nvidia |
Run pip install voicepad-core[gpu] |
Performance Benchmarks
RTX 3050 (4GB VRAM)
- Model: medium (recommended)
- CPU time: ~8-10s per minute of audio
- GPU time: ~2-3s per minute of audio
- Speedup: 4-5x
GTX 1050 (2-4GB VRAM)
- Model: small (recommended for limited VRAM)
- CPU time: ~6-8s per minute of audio
- GPU time: ~2-3s per minute of audio
- Speedup: 3-4x
Key: Speedup depends on model size. Smaller models are faster on GPU; larger models benefit more but need more VRAM.
FAQ
Q: Do I need CUDA Toolkit installed globally?
A: No. The [gpu] extra provides everything in your venv.
Q: Will GPU support conflict with my system CUDA? A: No. Virtual environment isolation prevents conflicts.
Q: Can I use AMD GPUs? A: Not currently. faster-whisper only supports NVIDIA CUDA.
Q: What about Apple Silicon (M1/M2)? A: GPU acceleration not available, but CPU mode works well.
Q: How do I disable GPU support?
A: Set transcription_device: cpu in voicepad.yaml, or uninstall: pip uninstall nvidia-cublas-cu12 nvidia-cudnn-cu12
Q: Can I use uvx voicepad with GPU?
A: Yes! uvx --with voicepad-core[gpu] voicepad record start
See Quick Start for installation or Advanced Setup for detailed requirements.