Crash when using CUDA on Windows

[no1d]

I'm experimenting around with the server on Windows and it works fine using CPU.
But with CUDA it crashes after a client disconnects. Seems like the speech-to-text thread doesn't die properly.
In Event Viewer I see this:

Faulting application name: python.exe, version: 3.10.11150.1013, time stamp: 0x642cc427
Faulting module name: ucrtbase.dll, version: 10.0.26100.1882, time stamp: 0x52db7a2d
Exception code: 0xc0000409
Fault offset: 0x00000000000a502e

or

Faulting application name: python.exe, version: 3.10.11150.1013, time stamp: 0x642cc427
Faulting module name: KERNELBASE.dll, version: 10.0.26100.2454, time stamp: 0x398a1cce
Exception code: 0xe06d7363
Fault offset: 0x00000000000c837a

For CUDA I've used

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Win11, 4070 SUPER, latest drivers