How to make torchaudio select the customized compiled ffmpeg

1

I installed pytorch using miniconda by

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

Now I want to use the GPU video decoder/encoder of torchaudio.

I compiled and installed a ffmpeg with cuda support following the guide in Enabling GPU video decoder/encoder — Torchaudio 2.0.1 documentation (pytorch.org)

However, after installation, torchaudio cannot pick the correct ffmpeg outside miniconda since ffmpeg is installed by conda as a dependency while installing pytorch.

How can I fix this problem?

2

You might need to build torchaudio from source and let it pick your custom library during the build.

3

Ended up with compiling and installing everything (includingffmpeg and cuda) from source into conda env. torchaudio does not need re-compilation and it picks the correct ffmpeg then.