Problem on compatibility with torch.no_grad() on Ubuntu GPU

综合讨论区
#1

If I comment out the torch.no_grad() in the testing, everything works fine, but it does not when the torch.no_grad() exists.
In addition, when the error happens, Volatile GPU-Util is 100%.
To be honest, I cannot reproduce this error in a simple file. The error information is below:

python(PyEval_EvalCode+0x1c) [0x564c01d613ec]
python(+0x22f874) [0x564c01e79874]
python(PyRun_FileExFlags+0xa1) [0x564c01e83b81]
python(PyRun_SimpleFileExFlags+0x1c3) [0x564c01e83d73]
python(+0x23ae5f) [0x564c01e84e5f]
python(_Py_UnixMain+0x3c) [0x564c01e84f7c]
/lib/x86_64-linux-gnu/libc.so.6: __libc_start_main
python(+0x1e0122) [0x564c01e2a122]
#2

Thanks for reporting this issue. It’s hard to judge what exactly happened according to the log. If you can share with me your code to reproduce (it doesn’t have to be short, if it’s hard to reproduce in a single file), that would be great!

#3

Update: the issue is fixed after upgrading PyTorch from 1.0.0 to 1.2.0.