Pearu Peterson
Pearu Peterson
Whatever are the reasons for scipy to use float64 internally (one practical reason could be that there are no float32 implementation available for scipy, for instance), evaluating functions using float32...
Heads up: https://github.com/jax-ml/jax/pull/27107 fixes all issues reported here.
Closing as currently irrelevant (no src/torchaudio/backend/no_backend.py in main).
Closing as currently irrelevant (no test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py in main)
`CODEOWNERS` does not currently have a mapping for sparse codes.
I think @nikitaved idea in https://github.com/pytorch/pytorch/issues/169954#issuecomment-3641746716 could be a realistic compromise: when invariants checks are not enabled explicitly (globally via `check_sparse_tensor_invariants` or locally via `check_invariants=` argument), there will be one-time...
Makes sense. Would you like to create a PR fixing this issue?
@anferico, the change looks good to me. Could you implement the corresponding sanity tests in this PR?
> The failure occurs because the new `torchcodec` backend, now a default dependency, is linked against `libpythonX.Y.so`. Can you raise this issue in `torchcodec`?
soundfile backend was removed in https://github.com/pytorch/audio/pull/4041 per https://github.com/pytorch/audio/issues/3902, so this issue is not actionable. The suggested alternative sounds like a reasonable approach.