base<\/code> environment.<\/p>\n\n\n\nTo activate the base<\/code> anaconda<\/code> environment,<\/p>\n\n\n\n$ module load anaconda\n$ conda activate\n(base) $ python\nPython 3.11.4\n>> ^D\n(base) $<\/code><\/pre>\n\n\n\nFirst deactivate the current conda<\/code> environment before unloading anaconda<\/code><\/p>\n\n\n\n(base) $ conda deactivate\n$ module unload anaconda<\/code><\/pre>\n\n\n\nBuilding CUDA code<\/h3>\n\n\n\n
Our recommendation for building CUDA code on a system without the cuda driver installed is to link against the “stub” libraries.
The Linux linker (ld) and dynamic loader (ld.so) handle libcuda.so and libcuda.so.1 based on the standard ELF shared library versioning conventions, with a specific, unconventional twist provided by NVIDIA. <\/p>\n\n\n\n
\n- libcuda.so (Link Time): Typically a symbolic link to libcuda.so.1 or a “stub” library located in \/usr\/local\/cuda\/lib64\/stubs\/ used only during compile-time linking to satisfy ld.<\/li>\n\n\n\n
- libcuda.so.1 (Run Time): The actual SONAME expected by the dynamic linker at runtime, which points to the real driver file (e.g., libcuda.so.535.104.05).\u00a0<\/li>\n\n\n\n
- The stub is designed so that the dynamic linker will refuse to use it at runtime, forcing it to find the real driver-installed libcuda.so.1 instead.<\/li>\n\n\n\n
- At runtime, the dynamic loader ignores the file name used at link time and looks for the embedded SONAME (libcuda.so.1) in LD_LIBRARY_PATH<\/li>\n<\/ul>\n\n\n\n
The MILC Makefile needs to be changed: <\/p>\n\n\n\n
< LIBQUDA += -L${CUDA_HOME}\/lib64 -L${CUDA_MATH}\/lib64 -L${CUDA_COMP}\/lib -lcudart -lcuda -lcublas -lcufft -ldl\n> LIBQUDA += -L${CUDA_HOME}\/lib64 -L${CUDA_MATH}\/lib64 -L${CUDA_COMP}\/lib -L${CUDA_HOME}\/lib64\/stubs -lcudart -lcuda -lnvidia-ml -lcublas -lcufft -ldl<\/code><\/pre>\n\n\n\nTo prevent libstdc++ clashes, build all linked codes, e.g., qmp, qio, quda with the same toolchain.<\/p>\n\n\n\n
For GCC + OpenMPI, use modules<\/p>\n\n\n\n
$ module purge\n$ module load gompi ucx_cuda ucc_cuda<\/code><\/pre>\n\n\n\nFor Intel toolchain, use<\/p>\n\n\n\n
$ module purge\n$ module load intel<\/code><\/pre>\n\n\n\nThe above also loads a compatible version of gcc\/g++.<\/p>\n","protected":false},"excerpt":{"rendered":"
HPC toolchains on LQ cluster complex The HPC toolchains on the LQ complex were built using a combination of EasyBuild recipes in combination with additional hand-built packages. The EasyConfig files used to build the software for LQ are on gitHub. Under AlmaLinux 8 (el8) the deployments are found in the directory \/srv\/software\/el8\/x86_64. The subdirectory .\/eb… More »<\/a><\/p>\n","protected":false},"author":16,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":"[]"},"class_list":["post-421","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/pages\/421","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/comments?post=421"}],"version-history":[{"count":67,"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/pages\/421\/revisions"}],"predecessor-version":[{"id":8068,"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/pages\/421\/revisions\/8068"}],"wp:attachment":[{"href":"https:\/\/computing.fnal.gov\/lqcd\/wp-json\/wp\/v2\/media?parent=421"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}