Now I’m running Caffe2 deep learning framework and wanting to debug CUDA code. I can debug the CUDA code from cuda-gdb session. However, when I want to attach to a running Caffe2 application, it can’t work. The following is the output:
Attaching to a running CUDA process with software preemption enabled in the debugger is not supported
I read the cuda-gdb manual, which told me that it’s now possible to attach to a CUDA application that is already running.
The GPU for me is Tesla K80 and the version of driver is 375.26. And the operating system for me is CentOS-7.3.
May I mislead the meaning on the cuda-gdb manual?
Does anybody encounter the same problem? And can anybody help me?