Assertion failure during transpiling in demo_chameleon

[wmdi]

I run into the following issue in the catalyst cluster:

$ python demo/demo_chameleon.py 
========== Search Configuration ==========
  max num threadblock graph op: 9
  max num kernel_graph op: 7
  max num threadblock graphs: 1
  max num threadblock graph inputs: 3
  max num threadblock graph outputs: 2
  search_thread: 8
  imaps to explore:
  imap combs to explore:
  omaps to explore:
  grid dims to explore:
  block dims to explore:
  fmaps to explore:
  franges to explore:4 16 64 
[Search] States: 901, Random tests: 1, Valid mugraphs: 0, Time: 4.766257
[Search] First step finished. Time elapsed: 4.766711sec
[Search] States: 743301, Random tests: 4307, Valid mugraphs: 16, Time: 172.965338
[Search] Second step finished. Time elapsed: 172.990405sec
[Search] Total states explored: 743323
[Search] Random tests performed: 4307
[Serach] Valid kernel graphs explored: 16
Transpiling muGraph 0...
Profiling muGraph 0 performance (ms) = 0.06813900756835937
Transpiling muGraph 1...
Profiling muGraph 1 performance (ms) = 0.06811238098144531
Transpiling muGraph 2...
python: /home/mengdiwu/mirage/src/threadblock/element_binary.cc:51: mirage::threadblock::STensor mirage::threadblock::Graph::elementbinary(const mirage::threadblock::STensor&, const mirage::threadblock::STensor&, mirage::type::TBOperatorType): Assertion `op != nullptr' failed.
Aborted (core dumped)

A potential reason is the inconsistency between memory checking in search and in transpiler: we modify the graph before transpilation, which may increase the memory usage, and then cause oom issue.