If you fail to build wheels on compiling c++, try this.

[lihe50hz]

1. Find this TransformerEngine/transformer_engine/pytorch/csrc/userbuffers/userbuffers.cu.
2. Change all half in this file to TransformerEngine_half or some other string you like which won't be used that often.

The error comes from #define te_half nv_bfloat16 at line 12 in the file.

[timmoon10]

This hack shouldn't be necessary with TE 1.9 or newer: #949

I've also gone ahead and cherry-picked that commit into TE 1.8: 5506efc