1.2

Fix fastattn bug.

Tensorflow 2.4

update for tf2.4 (#908)

* update for tf2.4

* fix mixed precision with recompute gradient

* update README

* fix multi gpus training

* update README

* fix LossScaleOptimizer bug

* disable steps_per_execution in default

* split all reduce

Stable version

Fix ema  (#883)

* fix ema

* update README