Open yondonfu opened 3 years ago
https://github.com/livepeer/go-livepeer/pull/1979/commits/c17d16e113775f0957bcaefb41a65b66dd756ea6 should fix this in the short-term by hardcoding tensorflow models to always run on whatever the system calls GPU 0. This will be consistent with what we did prior to #1979.
The way we currently configure tensorflow devices in ffmpeg, whatever device we pass through visible_device_list
should be mapped to a virtual tensorflow device "0".
According to the linked issue in the logs https://github.com/tensorflow/tensorflow/issues/19083 this is supported now, so we should keep this issue open to investigate why it failed for us here.
TODOs
This occurred when using the branch for https://github.com/livepeer/go-livepeer/pull/1979.