I mean, that's kind of the point. These are things you can make work with these frameworks, you just have to opt out of all optimizers and JIT compilers to make

I mean, that's kind of the point. These are things you can make work with these frameworks, you just have to opt out of all optimizers and JIT compilers to make it work (including TorchScript). Yes, there are some examples you can get away with that because the matmuls are big enough. No, that's not every model you could ever use, or even close to that. In fact, in the context of ODEs we found for example torchdiffeq is ~20x-5000x slower than it should be on simulating many scientific models, in large part because TorchScript optimizations are not able to be enabled on that code (https://gist.github.com/ChrisRackauckas/cc6ac746e2dfd285c28e0584a2bfd320 https://gist.github.com/ChrisRackauckas/cc6ac746e2dfd285c28e... , and it shows that enabling TorchScript does nothing, and you can trace it back to the dynamic constructs). This is thus a proof of concept of some ML algorithms that are then not handled by the optimizers and require the fallbacks to the slower setups.