Failed to compile torchvision for ROCm as documented in pytorch.org

🐛 Bug

Failed to compile torchvision for ROCm as documented in pytorch.org/get-started

To Reproduce

Steps to reproduce the behavior:

as in: https://pytorch.org/get-started/locally/

1. python -m venv ptamd; source ptamd/bin/activate
2. pip install torch -f https://download.pytorch.org/whl/rocm4.0.1/torch_stable.html
3. pip install ninja && pip install 'git+https://github.com/pytorch/vision.git@v0.9.1'

same with v0.9.0

Error

ptamd/lib/python3.8/site-packages/torch/include/c10/util/complex.h:9:10: fatal error: 'thrust/complex.h' file not found
#include <thrust/complex.h>
^~~~~~~~~~~~~~~~~~
1 error generated when compiling for gfx803.

Environment

PyTorch version: 1.8.1+rocm4.0.1
Is debug build: False
ROCM used to build PyTorch: 4.0.20496-4f163c68

OS: CentOS Linux 8 (x86_64)
GCC version: (GCC) 8.3.1 20191121 (Red Hat 8.3.1-5)   # same on GCC 10

Python version: 3.8 (64-bit runtime)
Is CUDA available: True
GPU models and configuration: Vega 20
HIP runtime version: 3.21.2
MIOpen runtime version: 2.9.0

Versions of relevant libraries:
[pip3] numpy==1.20.2
[pip3] torch==1.8.1+rocm4.0.1

Start with the documented installation commands in the issue, using the listed PyTorch 1.8.1+rocm4.0.1, torchvision v0.9.0 or v0.9.1, and CentOS environment. Investigate the compilation failure at torch/include/c10/util/complex.h:9, where thrust/complex.h is missing. Done means the documented torchvision installation completes successfully for the reported ROCm setup.