pytorch

3c6d9fd1 - Eagerly populate python_error::what() when TORCH_SHOW_CPP_STACKTRACES=1 (#65376)

Eagerly populate python_error::what() when TORCH_SHOW_CPP_STACKTRACES=1 (#65376) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/65376 Let's suppose there's a bug in PyTorch and python_error gets thrown and never gets caught. Typically, you'll get a very useless error message like this: ``` terminate called after throwing an instance of 'python_error' what(): Aborted (core dumped) ``` Now, you'll get: ``` what(): unknown Python error (for more information, try rerunning with TORCH_SHOW_CPP_STACKTRACES=1) ``` and with TORCH_SHOW_CPP_STACKTRACES=1 you'll get: ``` what(): error message from Python object ``` If we're OK with making Python exceptions go even slower, we could eagerly populate unconditionally. I'm also not so happy we don't get a Python backtrace or the Python error name, that's worth improving (this is a minimal diff to get the discussion going.) Signed-off-by: Edward Z. Yang <ezyang@fb.com> Test Plan: Imported from OSS Reviewed By: albanD Differential Revision: D31067632 Pulled By: ezyang fbshipit-source-id: 9cfda47cafb349ee3d6853cdfb0f319073b87bff