Very uninformative stack-trace when exception thrown in custom autograd backward #41659
Description
I had message-less assert in my custom autograd function backward. Note absence of the function name or class name. This happens with conda python. I think this is standard Python problem: allow_unreachable=True) # allow_unreachable flag, but maybe this is related to Python/C++ interaction?
Traceback (most recent call last):
File "benchmark.py", line 103, in <module>
y.sum().backward()
File "/miniconda/lib/python3.7/site-packages/torch/tensor.py", line 198, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/miniconda/lib/python3.7/site-packages/torch/autograd/__init__.py", line 100, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError
In fact this is an AssertionError
cc @ezyang @ssnl @albanD @zou3519 @gqchen @yf225 @glaringlee