From v0.3 to v0.4

No more CUDA-specific images

All users should switch to the latest image tag going forward. No new versions will be published to the cuda-latest tag. You can now more easily switch your transcription target device with the transcription engine settings.

The way this works is that the first time the container runs a transcription job with a given transcription engine/build target, it will install any necessary dependencies (e.g. OpenBLAS or the CUDA toolkit) and build whisper.cpp from source for that target. This may take several minutes, but it will only happen the first time running a transcription job after an upgrade.

No more CUDA-specific environment variables

Storyteller now supports CUDA devices for transcription with no environment variable configuration. That means you no longer have to restart your container in order to try out new configurations.

Note: The Storyteller container defaults to setting NVIDIA_VISIBLE_DEVICES=all and NVIDIA_DRIVER_CAPABILITIES=compute,utility. If you have multiple NVIDIA GPUs and wish to only expose one or some of them to the Storyteller container, you may want to set NVIDIA_VISIBLE_DEVICES to a different value in your compose file.