Defined in File inference.h
fuse SharedDeviceTensor oprs
This would treat all SharedDeviceTensor operators as constant, and replace oprs that only depend on them by the evaluated value at compile time.
Usually this pass is used after ParamRedistributePass.
set the limit for max param size growth due to merging
Param size may grow if param fusing causes low-rank result (i.e. by broadcasting). Size growth is defined to be the difference between new param size and max size of source oprs that it depends on.
This limit is given in bytes