Seg fault with test_rnn_retain_variables on ppc64le

[deepali-c]

🐛 Bug

Observed segfault with the test case test_rnn_retain_variables:

Other tests that crash are:

test_cuda_rnn_fused
test_rnn_initial_hidden_state
test_einsum

To Reproduce

The minimum to code to reproduce is:

import torch
import torch.nn as nn
device="cpu"
dtype=torch.double
rnn = nn.GRU(10, 20, num_layers=2).to(device,dtype)
input = torch.randn(5, 6, 10, device=device, dtype=dtype, requires_grad=True)
output = rnn(input)

Other observations are:
a. The above code works for device="cuda"
b. The above code works for dtype=torch.float
c. The above code works if the following is used for input:
input = torch.randn(3, 3, 10, device=device, dtype=dtype, requires_grad=True)

Expected behavior

The tests should pass.

Environment

PyTorch version: 1.0.0a0+7998997 (with some local changes)
Is debug build: No
CUDA used to build PyTorch: 10.0.130

OS: Red Hat Enterprise Linux Server 7.6 (Maipo)
GCC version: (GCC) 4.8.5 20150623 (Red Hat 4.8.5-36)
CMake version: Could not collect

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 10.0.130
GPU models and configuration:
GPU 0: Tesla V100-SXM2-16GB
GPU 1: Tesla V100-SXM2-16GB
GPU 2: Tesla V100-SXM2-16GB
GPU 3: Tesla V100-SXM2-16GB

Nvidia driver version: 410.72
cuDNN version: Could not collect

Additional context

[ezyang]

Possibly related: the test occasionally times out on non-PPC: https://circleci.com/gh/pytorch/pytorch/2000995?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link