grammar fix

Author: ckkuang

rfcs/20200928-sharded-variable.md

@@ -36,7 +36,7 @@ From our experiences, parameter servers often become the bottleneck of the train
 
 TF 1.x supports [PartitionedVariable](https://www.tensorflow.org/api_docs/python/tf/compat/v1/get_variable) and this feature is currently missing in TF 2.x. We'd like to add some level of support along with the new `ParameterServerStrategy` proposed in the aforementioned single-client parameter server training [RFC](https://github.com/tensorflow/community/pull/214).
 
-Note that previously a [RFC](https://github.com/tensorflow/community/blob/05e6343409a03d203e676b8c37a89370e6e67dae/rfcs/20190116-embedding-partitioned-variable.md) was published regarding sharded variable for TF 2.x. As things have evolved, especially with the proposal of the new `ParameterServerStrategy`, we feel it is necessary to write a new RFC to accommodate all the moving pieces. The notable differences between this RFC and the old RFC are:
+Note that previously an [RFC](https://github.com/tensorflow/community/blob/05e6343409a03d203e676b8c37a89370e6e67dae/rfcs/20190116-embedding-partitioned-variable.md) was published regarding sharded variable for TF 2.x. As things have evolved, especially with the proposal of the new `ParameterServerStrategy`, we feel it is necessary to write a new RFC to accommodate all the moving pieces. The notable differences between this RFC and the old RFC are:
 
 1. **APIs**. This RFC proposes to use variable creator scope (namely, strategy.scope()) to control the creation of sharded variables, while the old RFC proposed to add a method to strategy: `strategy.create_sharded_variable`. We believe that variable creator scope approach is less intrusive to layers.
 2. **Layer partition**. The old RPC focused on partitioning embedding layers, while this RPC proposes to support partitioning any type of layers. We have seen performance improvements for certain models when dense layers are partitioned (b/160977486).