pytorch

76697a3b - Enabled bfloat16 dtype on CUDA

Enabled bfloat16 dtype on CUDA Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/26407 Differential Revision: D17450502 Pulled By: izdeby fbshipit-source-id: 0a5acc5fe1b1555c61ebe038aee9eaaae9dac228