OOM when using VAD

[Troffifi]

Hi, does somebody else experience issues with memory consumption when transcribing audio files containing a lot of speech (~ 4 hours long)? I am running the latest version of faster-whisper in a Kubernetes pod on a g4dn AWS instance. The server has 4 cores, 1 GPU, and 16GB RAM, but the pod is limited to 2 cores. The base image is pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime and as per this pinned issue the installed versions should be compatible:

• python 3.11
• torch 2.5.1+cu124
• ctranslate2 4.5.0
• cuda 12.4
• cudnn 9.1.0.7

The process gets killed during the transcription phase when VAD is enabled. I tried the solution described here, but it doesn't help. See the logs attached. Anyone has any idea what could be the cause of the OOM?

libraries.txt
logs on sigkill.txt

[Purfview]

Try this fix -> #1198

[Troffifi]

I still have to run some tests, but I tried it and it looks promising. The latest commit has fixed it. Thank you @Purfview! I will let you know if the issue persists.