• OS: Linux
• PyTorch version: 0.4.0a0+eeb90d9
• Python version: 3.5

The following segfaults:

import torch
import numpy as np
data = np.array([1.0])
torch.tensor(data, dtype=torch.float32)
torch.tensor(data, dtype=torch.cuda.float32)

The first tensor call with dtype=torch.float32 seems necessary to trigger the crash on the second line.

(Reported by @Balandat)

cc @gchanan