3DFlex Dependencies; Building pycuda

Testing the 3D flex refine install on a system which has just finished a big run (and project is stored safely…) Ubuntu 22.04, with CUDA 11.8.

Update to 4.1.0 went smoothly. 3D flex refine install went smoothly until it hit PyCUDA, then did the now recognisable PyCUDA build fail.

During the 3D flex install, it seems to be downloading a mix of CUDA 11.7 and CUDA 12. CUDA 12 currently causes a world of pain for all of the 3DEM programs I’ve actually tried it with.

The PyCUDA build references very old compute targets, which might have been deprecated in CUDA 12?

Would anyone else please check whether they see cuda-cudart-12, cuda-cudart-dev-12 etc in their install list?

Testing a manual PyCUDA install now, but unlike getting it working for Arch, this is just sitting on “Solving environment” indefinitely.

Can provide logs if necessary. Updates if the manual PyCUDA install works.

Update: manually installing PyCUDA eventually throws dozens of compatibility errors (but anaconda took about three hours to work it all out). Will revert to non-3D Flex install for now.

1 Like