* Optimize the perf of top_k when k is too large

* fix rcom compile

* fix

* only compile in cuda

* fix log info