Repository navigation
Container fails to start on linux/arm64: basic-pitch has no inference backend #12
Description
Activity
Correction while working on this in #18: I wrote above that the
onnxextra would also shrink the image. That is wrong.The extra adds a backend but does not remove
tensorflow, which basic-pitch depends on unconditionally for non-Darwin platforms. Measured on arm64, the image grew from 1.4 GB to 1.46 GB.Dropping TensorFlow in favour of ONNX Runtime alone would genuinely shrink the amd64 image, but it changes which backend runs in production and would need a transcription-accuracy check, so it is out of scope for the startup fix. Worth its own issue if image size matters.
On Apple Silicon (
linux/arm64) the built image never starts. Shimmy waits 15 s for the worker, the worker exits 1, and the container dies.Cause: basic-pitch picks a backend at import and finds none.
poetry.lockpinstensorflowfor Linux, but the aarch64tensorflowwheel on PyPI is an empty shim that requirestensorflow-cpu-aws, which Poetry never resolved into the lock. Importingtensorflowinside the arm64 container raisesModuleNotFoundError.CI and the AWS build are x86_64, so they get the real wheel and pass. This only bites local development on arm64 machines.
Fix: add the
onnxextra to basic-pitch inpyproject.toml:onnxruntimehas aarch64 wheels and is far smaller than TensorFlow, so this also shrinks the image. Verified locally: withonnxruntimeinstalled the worker imports in 1.0 s and transcription works.