Ye Luo

could you suggest a reviewer?

> Yes, already flagged @anbenali for this. Oh thanks. I missed it.

I remember some discussion long time ago that twistnum is error-prone and explicit twist coordinates are preferred. Can nexus move away from twistnum?

@ppanchad-amd thank you for the message. I understood a clean uninstall allows me to avoid hassles. However, the issue I raised here is to see if direct upgrading can be...

Just noticed that e4s was doing some nightly tests. https://gitlab.e4s.io/uo-public/qmcpack/-/pipelines Although they used a node with a100 but no GPU feature is turned on.

Once https://github.com/spack/spack/pull/36675 gets in we should be able to work on 1. Updating +cuda variant if needed 2. add +rocm variant with the rocm package 3. add +offload variant

~~https://github.com/spack/spack/pull/36675~~ has been merged. If we enable offload features of qmcpack in spack, we can test the integration. When using llvm compilers, add `+libomptarget` requirement.

The gauging parameter is `walker_per_rank` since the number of maps is proportional to it. The above 1-4 MPI rank tests has 2048 walkers per rank. Only small problem sizes can...

Update. I can reproduce the issue without MPI. I built qmcpack without MPI and launch it 4 times manual. ``` NUM=0; CUDA_VISIBLE_DEVICES=0,1,2,3 numactl -N $NUM $my_path/qmcpack --enable-timers=fine $file_prefix.xml >& $NUM.log...

FYI. I posted the issue to NVIDIA forum with a minimal reproducer https://forums.developer.nvidia.com/t/cudahostregister-returns-cudaerrormemoryallocation-out-of-memory-in-runs-on-a-multi-gpu-node/337793/6?u=xw111luoye