Runtimeerror cudnn_status_execution_failed( cudnn_batch_norm_backward allow unreachable flaf

I updated my torch to PyTorch 1.6 as you said but the error stilled to appear. Finally I found the error is about one equation divided by zero :smile: :smile:. It is not about the memory size. Thank you alot for your patient and kindness

Do you have by any chance a code snippet to reproduce this issue?
A division by zero shouldn’t yield a cudnn error, which seems to be quite misleading.

mmm you are right. I don’t have a good explanation for this. but I modify this equation and the script done well. That’s what happened

here is the same error again. you are right. i am really confused and dont know what should i do

Rerun the script with the launching command in PyTorch 1.7 and post the stack trace.
If you get stuck, post an executable code snippet, so that we can debug it.