Strange problem using DataParallel

Since pytorch 0.1.11, I had a strange error using DataParallel.

I have some code which has no error using a single GPU.

However, using multiple GPUs, I got ‘out of memory error’ even with the same batch size.

(Note that this code works without any error on pytorch 0.1.10)

Does anyone experience a similar problem since pytorch 0.1.11 ?

Now, I am trying to write the small snippet reproducing this problem.

We’ve run many programs fine with v0.1.11.
I’m happy to help investigate further if you get me a snippet.