🐛 Bug

Shuffling a huge tensor by reindexing into it results in all of the values past a certain point being set to 0 when the reindexing is done in GPU memory.

@colesbury This likely relates to #20562 as the issue is also with reindexing and also shows up at the same point in the tensor. Hopefully this helps with narrowing the other issue.

I initially thought this was limited to longtensors, but if you use a floattensor with 2x values you run into the exact same issue in the exact same place.

To Reproduce

Steps to reproduce the behavior:

1. Create a huge longtensor (>536870700 values) of random values. I went with 15Mx45
2. Load it into gpu memory
3. Reindex to shuffle the values

Data past ~536870700 will be all 0. The data is zero'd mid tensor.

Here's a notebook that reproduces the issue:
https://github.com/EvenOldridge/HugeTensor/blob/master/Huge%20Tensor%20Bug%20-%20Data%20Loss.ipynb

Expected behavior

Data should not be lost when reindexing

Environment

• PyTorch Version (e.g., 1.0): 1.01
• OS (e.g., Linux): Linux (Ubuntu 16.04)
• How you installed PyTorch (conda, pip, source): pip
• Python version: 3.6
• CUDA/cuDNN version: 10.0
• GPU models and configuration: TeslaV100 32 Gig (on a DGX-1)