Preserve python backtrace in autograd engine errors. #43608
Conversation
This PR attempts to address #42560. We add a "Python Traceback" component to the error message recording the original location from where the exception was thrown. This approach isn't ideal since we're extending the exception message itself. The alternative I considered was extending our Future to record this information in addition to just the message. The tricky part about this was avoiding python dependency. Ideally, I'd like to avoid a python dependency in our Future class. Even if we do avoid the python depedency in our base class by creating a subclass for python, we would still end up having a python dependency in things like autograd engine's GraphTask. For the example in #42560, the exception trace would now look like: ``` > Traceback (most recent call last): > File "test_autograd.py", line 6914, in test_preserve_backtrace > Foo.apply(t).sum().backward() > File "torch/tensor.py", line 214, in backward > torch.autograd.backward(self, gradient, retain_graph, create_graph) > File "autograd/__init__.py", line 127, in backward > allow_unreachable=True) # allow_unreachable flag > RuntimeError: something > Python Traceback (most recent call last): > File "torch/autograd/function.py", line 87, in apply > return self._forward_cls.backward(self, *args) > File "test_autograd.py", line 6909, in backward > raise ValueError("something") ``` Differential Revision: [D23337371](https://our.internmc.facebook.com/intern/diff/D23337371/) [ghstack-poisoned]
This PR attempts to address #42560. We add a "Python Traceback" component to the error message recording the original location from where the exception was thrown. This approach isn't ideal since we're extending the exception message itself. The alternative I considered was extending our Future to record this information in addition to just the message. The tricky part about this was avoiding python dependency. Ideally, I'd like to avoid a python dependency in our Future class. Even if we do avoid the python depedency in our base class by creating a subclass for python, we would still end up having a python dependency in things like autograd engine's GraphTask. For the example in #42560, the exception trace would now look like: ``` > Traceback (most recent call last): > File "test_autograd.py", line 6914, in test_preserve_backtrace > Foo.apply(t).sum().backward() > File "torch/tensor.py", line 214, in backward > torch.autograd.backward(self, gradient, retain_graph, create_graph) > File "autograd/__init__.py", line 127, in backward > allow_unreachable=True) # allow_unreachable flag > RuntimeError: something > Python Traceback (most recent call last): > File "torch/autograd/function.py", line 87, in apply > return self._forward_cls.backward(self, *args) > File "test_autograd.py", line 6909, in backward > raise ValueError("something") ``` Differential Revision: [D23337371](https://our.internmc.facebook.com/intern/diff/D23337371/) ghstack-source-id: 110722744 Pull Request resolved: #43608
|
Hi, I don't think that match the issue. For example, the |
As I mentioned in the PR summary, to propagate the original exception we need to store this information as part of the Future which would incur a python dependency: "This approach isn't ideal since we're extending the exception message itself. |
What I had in mind is to save it as a pytorch/torch/csrc/autograd/engine.cpp Line 391 in f531815 |
We probably need to save the |
|
Abandoning in favor of #43684. |
Stack from ghstack:
This PR attempts to address #42560. We add a "Python Traceback"
component to the error message recording the original location from where the
exception was thrown.
This approach isn't ideal since we're extending the exception message itself.
The alternative I considered was extending our Future to record this
information in addition to just the message. The tricky part about this was
avoiding python dependency. Ideally, I'd like to avoid a python dependency in
our Future class. Even if we do avoid the python depedency in our base class by
creating a subclass for python, we would still end up having a python
dependency in things like autograd engine's GraphTask.
For the example in #42560, the exception trace would now look like:
Differential Revision: D23337371