Added better error messages for type validators (stack trace)

[cptspacemanspiff]

Summary

As I have been using executorch, I have run into issues with dtypes conflict during export, I end up with dtype mismatch during the to_edge process in exir. Unfortunately the current error message is less than clear, this change adds the stack trace from the nodes that have the issues, making it easier to find in your code where the type mismatch is.

Original Error Message:

    raise SpecViolationError(
torch._export.verifier.SpecViolationError: These operators are taking Tensor inputs with mismatched dtypes: defaultdict(<class 'dict'>, {<EdgeOpOverload: aten.cat.default>: schema = aten::cat(Tensor[] tensors, int dim=0) -> Tensor: {'tensors': torch.int32, '__ret_0': torch.int64}})Please make sure the dtypes of the Tensor inputs are the same as the dtypes of the corresponding 

New Error Message:

    raise SpecViolationError(
torch._export.verifier.SpecViolationError: These operators are taking Tensor inputs with mismatched dtypes:

Operator: <EdgeOpOverload: aten.cat.default>: schema = aten::cat(Tensor[] tensors, int dim=0) -> Tensor with args: {'tensors': torch.int32, '__ret_0': torch.int64}
stack trace:   File "/home/nlong/execu-tools/python/execu_tools/encoder_decoder_export.py", line 280, in reset_encode_prefill
    new_tokens = torch.cat((past_decoder_outputs, next_tokens), dim=1)

Please make sure the dtypes of the Tensor inputs are the same as the dtypes of the corresponding outputs.

Changes:

Internally the validator is collecting the errors as they come up into a list, then later if that list is not empty an exception is made.

The only change is that during the collection process I also collect self.node, which happens to be the current node we are evaluating, and add that to the list (as the second item of a tuple). Then when raising the exception we have access to not just the op/args but also the node that was being evaluated.

Makes tracking these down much easier.

Test plan

No testing, aside from being able to use it to fix my problems, it should not have any issues.
self.node is already used to access the "meta" field, needed for the arg validation.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/7999

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 8 Pending

As of commit 777c710 with merge base 9bd18f6 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[swolchok]

LGTM, thanks!

[swolchok]

looks like this broke test_edge_dialect_fails in the linux unittest, please fix and ping me for re-review

[cptspacemanspiff]

@swolchok fixed the unit test (it now checks the contents of the first element of the new pair.)