Does PyTorch guarantee backwards compatibility between minor releases (ie., 1.X->1.Y for all Y > X) for models/operator behaviors? In other words, would all the models defined and saved using an older 1.X version be loadable and runnable, with same behavior, using latest 1.Y version? Or, is this best effort which implies that some models will fail with upgrade of PyTorch version from 1.X->1.Y
Similarly, does PyTorch JIT guarantee backwards-compatibility for traced/scripted models between releases?
I didn’t find any documentation on backwards compatibility guarantee and I see backwards incompatible changes being noted in the release notes of each version release. Could someone help answer this or point to a document?