The main goal of the PR is to increase the minimum pytorch version to support 2.0.x.
However, I ran into this bug https://github.com/pytorch/pytorch/issues/45208 so made a quick fix for it, and slightly increased the tolerance on the allclose check we do on exports. The diff in audio.py looks a bit off, but I just added a torch.inference_mode() call at the top.
The main goal of the PR is to increase the minimum pytorch version to support 2.0.x.
However, I ran into this bug https://github.com/pytorch/pytorch/issues/45208 so made a quick fix for it, and slightly increased the tolerance on the
allclose
check we do on exports. The diff inaudio.py
looks a bit off, but I just added atorch.inference_mode()
call at the top.