Sergii Dymchenko
Sergii Dymchenko
> BTW, is there any public logs that I can see to get the error info, so that I can deal with the problem. @wangxiyuan Unfortunately no, there is no...
> OK. I guess there are some tests/functions/tools inner Meta are still using the `_export` API here? So maybe the next step is to wait for the cleanup action in...
> For this PR, the usages are in tests. So that’s fine And in user code outside of PyTorch, we want to change `torch.onnx._export` to `torch.onnx.export`, right?
> Probably yes, though this result in some flags not being publicly available. Is there a more detailed guidance?
> Probably yes, though this result in some flags not being publicly available. Found another discrepancy: `torch.onnx.export` doesn't return anything, but `_export` returns. So this kind of code doesn't work...
> did this PR pass the import test? @thiagocrepaldi no, `_export` is used in several places, I'm trying to figure out what to do with that.
`assigntome` doesn't work in this repo, just commenting on the issue is fine.
But please mention the specific items from the list to you want to do.
@sshkhr You don't need to ask. Just check that there is no overlap with what others already commented on.
Agree. It might be problematic to figure out statically when `a` in `.to(a)` is a tensor, and not a dtype or device. Do you have any real-world examples?