PR_CI(PADDLE) is not stable.

[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

[guomingz]

From this summary link, it looks the host yq01-sys-hic-p40-0089.yq01 has the issue on the PR_CI(PADDLE) testing, you may refer to this

[luotao1]

yq01-sys-hic-p40-0089.yq01 is not stable, and @tianshuo78520a is fixing now.

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

The reason is that this PR is rerun several times. Maybe this PR is rerun by others, please notify the waiting lists to check it.

[wojtuss]

Closed, due to INT8v1 having been removed.