Replace NcclReduce/Broadcast ops during graph optimization so that we can...
Replace NcclReduce/Broadcast ops during graph optimization so that we can generate gradients for Reduce/Broadcast. Changing _NcclBroadcastRecv shape input to int32 so that the corresponding Const op is outputting to HostMem. PiperOrigin-RevId: 172279684
Loading
Please sign in to comment