CUDA error: "Unknown error" (222)

[zcalamari]

Hello,

I am trying to get the GPU optimized beagle library to work with beast2 (Running on Ubuntu 24.04 with CUDA 12.8). I installed both beagle-lib and beast2 from source, and beast2 recognizes that there is a GPU available, but when I run beast, I get the following error:

"CUDA error: "Unknown error" (222) from file </home/zcalamari/beagle-lib/libhmsbeagle/GPU/GPUInterfaceCUDA.cpp>, line 284."

In other posts with this error, it looked like OpenCL might be interfering, but I build beagle without that option:

cmake -DCMAKE_INSTALL_PREFIX:PATH=$HOME -DBUILD_OPENCL=OFF ..

The installation didn't produce any errors, and I've tried this with the binary beast2 as well with the same error. Any further troubleshooting or things to try would be greatly appreciated!

Thanks,
Zac

[xji3]

Hi Zac,

Could you paste the whole screen output from your run to show more information? In addition, what GPU are you using, how much memory does it have, and how big are your data? Does beagle run for a tiny test case with GPU? Is it because your GPU does not have enough memory?

Thanks,
Xiang

[zcalamari]

Hi Xiang,

Thanks for the quick reply. I've attached a file with the full screen output. This is the output from beast -beagle_info:

1 : NVIDIA GeForce RTX 4080 SUPER
Global memory (MB): 16071
Clock speed (Ghz): 2.56
Number of cores: 10240
Flags: PRECISION_SINGLE PRECISION_DOUBLE COMPUTATION_SYNCH COMPUTATION_ASYNCH EIGEN_REAL EIGEN_COMPLEX SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS SCALERS_RAW SCALERS_LOG VECTOR_NONE THREADING_CPP THREADING_NONE PROCESSOR_GPU FRAMEWORK_CUDA PARALLELOPS_STREAMS PARALLELOPS_GRID

The xml file for this tree is 20 MB, and I get the same error when I try to run beast on the primate mtDNA data from the beast tutorials.

Best,
Zac

screen.txt

[xji3]

Hi Zac,

Sorry that I am not very familiar with BEAST 2. Is it possible to test your BEAGLE installation using BEAST X and send me the above information?

Thanks,
Xiang

[zcalamari]

Hi Xiang,

I tried aligning the CUDA library with the nvidia drivers (I had installed a more recent CUDA toolkit than what the driver listed as supported) and now the error is specifically an "out of memory" issue (CUDA error: "Out of memory" (2) from file </home/zcalamari/beagle-lib/libhmsbeagle/GPU/GPUInterfaceCUDA.cpp>, line 596.). Is there a setting to control this?

I cannot run the file in BEAST X as it produces an error about the xml file being poorly formatted due to being generated with a more recent version of BEAUti than is compatible.

Thanks,
Zac

[xji3]

Hi Zac,
It's good to know that missing error info could be caused by driver version mismatch. I will add this to the FAQs.

Unfortunately, for the "Out of memory" error, there isn't much to do unless either downsize the data set to run on the same card (so that smaller data set requires less memory) or getting more GPU cards to run the analysis. The latter requires you to separate your data into several smaller data sets (e.g., dividing the columns of the sequence alignment) such that each small data set can be run under one GPU card.

It seems that issue from BEAGLE side is cleared. I am closing this issue.

Thanks,
Xiang