Add support to AMD ROCm

[FindHao]

PyTorch could be easily installed with AMD ROCm. But torchbench has some limitations on ROCm envs.

• model installation
• model execution
• GPU memory measurement

[FindHao]

Models detectron2_* and opacus_cifar10 have installation errors. If we only want to measure execution time, torchbench works well on other models. Verified on MI210 and ROCm 5.2.0.

[FindHao]

Models detectron2_* and opacus_cifar10 have installation errors. If we only want to measure execution time, torchbench works well on other models. Verified on MI210 and ROCm 5.2.0.

For those models' installation, need to make sure rocm has its llvm installed.

[FindHao]

#1752 adds a userbenchmark for testing on ROCm.

[jinsong-mao]

how about running test_bency.py on rocm+amd?

[FindHao]

how about running test_bency.py on rocm+amd?

Are you asking about the current support status of rocm?