Shengbo Song

> I see the function > > https://github.com/tkestack/vcuda-controller/blob/5ec3fcbc58679bbbbd82ef9e21d647d8d7383876/src/hijack_call.c#L713-L731 > > so, should it be change to > > ```c > request_size = base_size * pAllocateArray->NumChannels * > pAllocateArray->Height * pAllocateArray->Width...

The gpu-manager version and log?

Please follow the FAQ, provides the application container log

You need following the FAQ, set environment first, then run your application, the expected log should have `/tmp/xxx` pattern

> I have followed the FAQ。 >  > I can't find any useful logs in this directory。Which log in this directory do you want。 > @mYmNeo 1. Export LOGGER_LEVEL...

> https://gist.github.com/xwttzz/1f4b3794a2fb19f430ebea828030d145 > The above link shows all logs after VCUDA is enabled。 > @mYmNeo The error shows you application missing library `ImportError: libnvinfer.so.7: cannot open shared object file: No...

How about re-pull the image thomassong/gpu-manager:1.1.5. We have fixed a recursive problem about vcuda.

> https://gist.github.com/xwttzz/1f4b3794a2fb19f430ebea828030d145 I think you should debug the coredump to find where it crash

Kubelet doesn't allow oversell device resource, how was your calculation?

gpu-manager will round up any memory size less than 256MB. Kubelet doesn't allow oversell, gpu-manager will not either.