Fuse RPC op to improve the performance of distributed training
Created by: Yancey1989
Currently, the RPC op like send_op
and recv_op
, they send/recv one variable for one RPC call, we can fuse these RPC calls to improve the performance.
tasks:
- modify RPC proto to support processing of multiple variables in an RPC message.
-
add a
FuseRPCPass
to fuse rpc op.