Created by: guomingz

Hi,

i found my PR 17130 met the PRE_CI(PADDLE) fail issue. After i checked the log, i found the error was not raised by code but the insufficient gpu memory. e.g [20:00:56] : [Step 1/1] unknown file: Failure [20:00:56] : [Step 1/1] C++ exception with description "cudaMemGetInfo failed in paddle::platform::GetMemoryUsage: out of memory at [/paddle/paddle/fluid/platform/gpu_info.cc:198]

But i got the pass after rerun testing.

Besides that, the PRE_CI(PADDLE) will rerun automatically, even the previous testing was passed.

Would you please give me some hints on this issue? @luotao1