🐛 Bug

When writing a new PR, CI reported RuntimeError: CUDA error: an illegal memory access was encountered for a test I added and all the subsequent tests in configuration pytorch_linux_xenial_cuda10_2_cudnn7_py3_gcc7_test.

To Reproduce

Steps to reproduce the behavior:

1. (Re-)trigger a CI run for [ignore] [WiP] reproducer #2 for 'RuntimeError: CUDA error: an illegal memory access was encountered' #36864

Expected behavior

No error, or a more descriptive error message.

Environment

CircleCI

Additional context

Shoutout to @pearu for narrowing it down and cc in case he has something to add.

#21819 has the same symptoms but according to #21819 (comment) it's a different issue.

cc @ezyang @gchanan @zou3519 @ngimel