🐛 Bug

MultiHeadAttention does not respect bias=False in the output projection. It seems that it was changed as part of a PR: eace053#diff-61bd5e4f390b965228ead6bb0efcad5b9b5299d833bcf5c2aa11ebaebb39ba29L794

Is this the intended behavior?

@ezyang

To Reproduce

Steps to reproduce the behavior:

1. Create a MultiHeadAttention module with torch.nn.MultiheadAttention(N, H, bias=False)
2. The module's out_proj field as a non-None, nonzero bias

Expected behavior

The bias in out_proj should be None.

Environment

• PyTorch Version (e.g., 1.0): 1.7.1 and latest master
• OS (e.g., Linux): All
• How you installed PyTorch (conda, pip, source): conda
• Build command you used (if compiling from source): N/A
• Python version: 3.8
• CUDA/cuDNN version: N/A
• GPU models and configuration: N/A
• Any other relevant information: N/A

Additional context

N/A

cc @ezyang @gchanan @zhangguanheng66