Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Enable torch.bfloat16 in AmpOptimWrapper and config file

Modification

New optional argument dtype in AmpOptimWrapper, can be str or torch.dtype

BC-breaking (Optional)

No

Use cases (Optional)

optim_wrapper = dict(
    type='AmpOptimWrapper',
    dtype='bfloat16',  # (None, float32, float16, bfloat16, ...)
    ....
)

Test results

mmcls::resnet50_8xb256-rsb-a3-100e_in1k

Hint: Amp with dtype=torch.bfloat16 works bad on convolutions, because it doesn't use CuDNN by default. Enable CuDNN version bfloat16 convolution by environment variable: TORCH_CUDNN_V8_API_ENABLED=1

mmdet::retinanet_r50_fpn_1x_coco

Failed due to torch.bfloat16 not supported by F.interpolate.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMCls.
4. The documentation has been modified accordingly, like docstring or example tutorials.