Yimin Tang
有个参数 "use_gpu=False"在 PolicyValueNet，可以确认一下是否打开。
Hi @whoenig , I think @john123741 is correct. It should use LB rather than cost in high-level search. https://tzin.bgu.ac.il/~felner/2014/cbseShort.pdf
When you are doing Lowlevel search ```c++ LowLevelSearch_t lowLevel(llenv, m_w); bool success = lowLevel.search(initialStates[i], newNode.solution[i]); ``` The newNode.solution[i] satisfy ```c++ newNode.solution[i].fmin
Yes, you are right. I didn't notice it is an unordered map.
@albanD @peterbell10
@peterbell10 @kulinseth Hi all, could you check this PR? Thanks.
@kulinseth Some tests failed. https://github.com/pytorch/pytorch/pull/80432#issuecomment-1170366368
@qqaatw Thanks! Just add the new commit.
@malfet @qqaatw I have tested the new patch. It doesn't work. Some .mm files use aten/src/ATen/native/DispatchStub.h ```c++ // NB: This macro must be used in an actual 'cu' file; if...
@peterbell10 @qqaatw @malfet I think this PR can be merged.