Forum Discussion
GPU implementation: Trying to solve a MILP problem with PDHG and GPU in an AWS machine.
I get this log: "GPU support is not available for the hybrid gradient algorithm in the deterministic concurrent LP solver. The CPU version will be used instead."
Am I doing something wrong or the PDHG algorithm is not supposed to work for MILPs.
These are my control settings:
DEFAULTALG = 4
OUTPUTLOG = 1
MPSNAMELENGTH = 64
CALLBACKFROMMAINTHREAD = 1
CONCURRENTTHREADS = 0
BARALG = 4
BARHGGPU = 1
GPUPLATFORM = 1
6 Replies
Hi Zejun,
DEFAULTALG in general only applies to LP resolves during cutting and in the tree. If you want to use PDHG for the root LP, you would need to set LPFLAGS=4.
Hello Tristan,
Thank you for your response.
I set the LPFLAGS = 4 and kept only the following controls:
BARALG = 4
LPFLAGS = 4
BARHGGPU = 1
Now this log appears :"Failed to load library xprscuda.?1163 Warning: GPU support requested but not available, not using GPU."
Do you have any indications to why this could happen?
There are generally two things this could imply, either the xprscuda library itself is not in a folder that is listed in the path or one of the libraries it depends on. Did you have a look at https://www.fico.com/fico-xpress-optimization/docs/latest/installguide/dhtml/chapinst1.html?scroll=secgpuinstall , in particular the required NVIDIA drivers and CUDA runtimes? Are you running this directly from the C library or, e.g., from python, where xprscuda would need to be installed separately?
I've checked the nvidia drives and cuda runtimes and the versions are eligible. Im running this in Julia so maybe thats the problem.
Can you double-check that both xprscuda and the required cuda framework libraries are in your PATH?
Yeah, i hadn't installed xprscuda seperately in julia. After installing the package it works fine.
Thanks,
