Merge pull request #367 from trz42/nessi-2023.06-cuTENSOR-2.0.1.2-CUDA-12.1.1

{2023.06}[system] cuTENSOR v2.0.1.2 with CUDA/12.1.1

Author: poksumdo

EESSI-install-software.sh

@@ -204,8 +204,10 @@ ${TOPDIR}/install_scripts.sh --prefix ${EESSI_PREFIX}
 # TODO: We should make a nice yaml and loop over all CUDA versions in that yaml to figure out what to install
 # Allow skipping CUDA SDK install in e.g. CI environments
 if [ -z "${skip_cuda_install}" ] || [ ! "${skip_cuda_install}" ]; then
-    ${EESSI_PREFIX}/scripts/gpu_support/nvidia/install_cuda_host_injections.sh -c 12.1.1 --accept-cuda-eula
-    ${EESSI_PREFIX}/scripts/gpu_support/nvidia/install_cuDNN_host_injections.sh -c 12.1.1 -d 8.9.2.26
+    ${EESSI_PREFIX}/scripts/gpu_support/nvidia/install_cuda_and_libraries.sh \
+        -e ${EESSI_PREFIX}/scripts/gpu_support/nvidia/eessi-2023.06-cuda-and-libraries.yml \
+        -t /tmp/temp \
+        --accept-cuda-eula
 else
     echo "Skipping installation of CUDA SDK and cu* libraries in host_injections, since the --skip-cuda-install flag was passed"
 fi

create_lmodsitepackage.py

@@ -105,45 +105,46 @@
 
 end
 
-
-local function eessi_cuda_enabled_load_hook(t)
+local function eessi_cuda_and_libraries_enabled_load_hook(t)
     local frameStk  = require("FrameStk"):singleton()
     local mt        = frameStk:mt()
     local simpleName = string.match(t.modFullName, "(.-)/")
-    -- If we try to load CUDA itself, check if the full CUDA SDK was installed on the host in host_injections. 
-    -- This is required for end users to build additional CUDA software. If the full SDK isn't present, refuse
-    -- to load the CUDA module and print an informative message on how to set up GPU support for NESSI
+    local packagesList = { ["CUDA"] = true, ["cuDNN"] = true, ["cuTENSOR"] = true }
+    -- If we try to load any of the modules in packagesList, we check if the
+    -- full package was installed on the host in host_injections.
+    -- This is required for end users to build additional software that depends
+    -- on the package. If the full SDK isn't present, refuse
+    -- to load the module and print an informative message on how to set up GPU support for NESSI
     local refer_to_docs = "For more information on how to do this, see https://www.eessi.io/docs/gpu/.\\n"
-    if simpleName == 'CUDA' then
+    if packagesList[simpleName] then
+        -- simpleName is a module in packagesList
         -- get the full host_injections path
         local hostInjections = string.gsub(os.getenv('EESSI_SOFTWARE_PATH') or "", 'versions', 'host_injections')
-        -- build final path where the CUDA software should be installed
-        local cudaEasyBuildDir = hostInjections .. "/software/" .. t.modFullName .. "/easybuild"
-        local cudaDirExists = isDir(cudaEasyBuildDir)
-        if not cudaDirExists then
+        -- build final path where the software should be installed
+        local packageEasyBuildDir = hostInjections .. "/software/" .. t.modFullName .. "/easybuild"
+        local packageDirExists = isDir(packageEasyBuildDir)
+        if not packageDirExists then
             local advice = "but while the module file exists, the actual software is not entirely shipped with NESSI "
-            advice = advice .. "due to licencing. You will need to install a full copy of the CUDA SDK where NESSI "
+            advice = advice .. "due to licencing. You will need to install a full copy of the " .. simpleName .. " package where NESSI "
             advice = advice .. "can find it.\\n"
             advice = advice .. refer_to_docs
             LmodError("\\nYou requested to load ", simpleName, " ", advice)
         end
     end
-    -- when loading CUDA enabled modules check if the necessary driver libraries are accessible to the NESSI linker,
+    -- when loading CUDA (and cu*) enabled modules check if the necessary driver libraries are accessible to the NESSI linker,
     -- otherwise, refuse to load the requested module and print error message
-    local checkGpu = mt:haveProperty(simpleName,"arch","gpu")
-    local overrideGpuCheck = os.getenv("EESSI_OVERRIDE_GPU_CHECK")
-    if checkGpu and (overrideGpuCheck == nil) then
+    local haveGpu = mt:haveProperty(simpleName,"arch","gpu")
+    if haveGpu then
         local arch = os.getenv("EESSI_CPU_FAMILY") or ""
-        local cudaVersionFile = "/cvmfs/pilot.nessi.no/host_injections/nvidia/" .. arch .. "/latest/cuda_version.txt"
-        local cudaDriverFile = "/cvmfs/pilot.nessi.no/host_injections/nvidia/" .. arch .. "/latest/libcuda.so"
+        local cvmfs_repo = os.getenv("EESSI_CVMFS_REPO") or ""
+        local cudaVersionFile = cvmfs_repo .. "/host_injections/nvidia/" .. arch .. "/latest/cuda_version.txt"
+        local cudaDriverFile = cvmfs_repo .. "/host_injections/nvidia/" .. arch .. "/latest/libcuda.so"
         local cudaDriverExists = isFile(cudaDriverFile)
         local singularityCudaExists = isFile("/.singularity.d/libs/libcuda.so")
         if not (cudaDriverExists or singularityCudaExists)  then
             local advice = "which relies on the CUDA runtime environment and driver libraries. "
             advice = advice .. "In order to be able to use the module, you will need "
-            advice = advice .. "to make sure NESSI can find the GPU driver libraries on your host system. You can "
-            advice = advice .. "override this check by setting the environment variable EESSI_OVERRIDE_GPU_CHECK but "
-            advice = advice .. "the loaded application will not be able to execute on your system.\\n"
+            advice = advice .. "to make sure NESSI can find the GPU driver libraries on your host system.\\n"
             advice = advice .. refer_to_docs
             LmodError("\\nYou requested to load ", simpleName, " ", advice)
         else
@@ -174,38 +175,13 @@
     end
 end
 
-local function eessi_cudnn_enabled_load_hook(t)
-    local frameStk  = require("FrameStk"):singleton()
-    local mt        = frameStk:mt()
-    local simpleName = string.match(t.modFullName, "(.-)/")
-    -- If we try to load cuDNN itself, check if the full cuDNN package was installed on the host in host_injections. 
-    -- This is required for end users to build additional cuDNN dependent software. If the full SDK isn't present, refuse
-    -- to load the cuDNN module and print an informative message on how to set up GPU support for NESSI
-    local refer_to_docs = "For more information on how to do this, see https://www.eessi.io/docs/gpu/.\\n"
-    if simpleName == 'cuDNN' then
-        -- get the full host_injections path
-        local hostInjections = string.gsub(os.getenv('EESSI_SOFTWARE_PATH') or "", 'versions', 'host_injections')
-        -- build final path where the cuDNN software should be installed
-        local cudnnEasyBuildDir = hostInjections .. "/software/" .. t.modFullName .. "/easybuild"
-        local cudnnDirExists = isDir(cudnnEasyBuildDir)
-        if not cudnnDirExists then
-            local advice = "but while the module file exists, the actual software is not entirely shipped with NESSI "
-            advice = advice .. "due to licencing. You will need to install a full copy of the cuDNN package where NESSI "
-            advice = advice .. "can find it.\\n"
-            advice = advice .. refer_to_docs
-            LmodError("\\nYou requested to load ", simpleName, " ", advice)
-        end
-    end
-end
-
 -- Combine both functions into a single one, as we can only register one function as load hook in lmod
 -- Also: make it non-local, so it can be imported and extended by other lmodrc files if needed
 function eessi_load_hook(t)
-    -- Only apply CUDA and cuDNN hooks if the loaded module is in the NESSI prefix
-    -- This avoids getting an Lmod Error when trying to load a CUDA or cuDNN module from a local software stack
+    -- Only apply CUDA and libraries hook if the loaded module is in the NESSI prefix
+    -- This avoids getting an Lmod Error when trying to load a CUDA or library module from a local software stack
     if from_eessi_prefix(t) then
-        eessi_cuda_enabled_load_hook(t)
-        eessi_cudnn_enabled_load_hook(t)
+        eessi_cuda_and_libraries_enabled_load_hook(t)
     end
 end
 

easystacks/pilot.nessi.no/2023.06/eessi-2023.06-eb-4.9.1-001-system.yml

@@ -3,4 +3,5 @@ easyconfigs:
       options:
         from-pr: 20299
   - EESSI-extend-2023.06-easybuild.eb
-# comment to trigger rebuild
+  - cuDNN-8.9.2.26-CUDA-12.1.1.eb
+  - cuTENSOR-2.0.1.2-CUDA-12.1.1.eb

easystacks/pilot.nessi.no/2023.06/eessi-2023.06-eb-4.9.1-2023a.yml

@@ -34,6 +34,5 @@ easyconfigs:
       # see https://github.com/easybuilders/easybuild-easyconfigs/pull/19451;
       options:
         from-pr: 19451
-  - cuDNN-8.9.2.26-CUDA-12.1.1.eb
   - OSU-Micro-Benchmarks-7.2-gompi-2023a-CUDA-12.1.1.eb
   - ABySS-2.3.7-foss-2023a.eb