Parallelize Transpose (#16854)

Commit

2 years ago

Parallelize Transpose (#16854) It gives up to 5.6% improvement for prompt and 2.3% improvement for token generation in LLaMA 7B case.