DEBUG only {2023.06}[2023a] PyTorch-Bundle v2.1.2 -- tweaked hooks employing LD_PRELOAD

bot/build.sh

@@ -159,7 +159,6 @@ echo "bot/build.sh: EESSI_OS_TYPE='${EESSI_OS_TYPE}'"
 declare -a COMMON_ARGS=()
 COMMON_ARGS+=("--verbose")
 COMMON_ARGS+=("--access" "rw")
-COMMON_ARGS+=("--mode" "run")
 [[ ! -z ${CONTAINER} ]] && COMMON_ARGS+=("--container" "${CONTAINER}")
 [[ ! -z ${HTTP_PROXY} ]] && COMMON_ARGS+=("--http-proxy" "${HTTP_PROXY}")
 [[ ! -z ${HTTPS_PROXY} ]] && COMMON_ARGS+=("--https-proxy" "${HTTPS_PROXY}")
@@ -179,6 +178,28 @@ fi
 [[ ! -z ${BUILD_LOGS_DIR} ]] && INSTALL_SCRIPT_ARGS+=("--build-logs-dir" "${BUILD_LOGS_DIR}")
 [[ ! -z ${SHARED_FS_PATH} ]] && INSTALL_SCRIPT_ARGS+=("--shared-fs-path" "${SHARED_FS_PATH}")
 
+# The following is only a proof-of-concept to replace the container's /bin/bash
+# with the bash from the compat layer.
+#
+# replace /bin/bash with bash from compat layer
+# - first obtain them
+# - then use them
+# EESSI_EPREFIX points to the compat layer
+# CMD='cp ${EESSI_EPREFIX}/bin/bash .'
+CMD="cp /cvmfs/software.eessi.io/versions/2023.06/compat/linux/${HOST_ARCH}/bin/bash ."
+get_bash_outerr=$(mktemp get_bash.outerr.XXXX)
+echo "Executing command to obtain bash executable from compat layer:"
+echo "./eessi_container.sh ${COMMON_ARGS[@]} --mode shell --storage ${STORAGE} <<< ${CMD} 2>&1 | tee -a ${get_bash_outerr}"
+./eessi_container.sh "${COMMON_ARGS[@]}" --mode shell --storage "${STORAGE}" <<< "${CMD}" 2>&1 | tee -a ${get_bash_outerr}
+
+compat_bash="${PWD}/bash"
+if [[ -z ${SINGULARITY_BIND} ]]; then
+    export SINGULARITY_BIND="${compat_bash}:/bin/bash"
+else
+    export SINGULARITY_BIND="${SINGULARITY_BIND},${compat_bash}:/bin/bash"
+fi
+
+COMMON_ARGS+=("--mode" "run")
 # determine if the removal step has to be run
 # assume there's only one diff file that corresponds to the PR patch file
 pr_diff=$(ls [0-9]*.diff | head -1)

easystacks/software.eessi.io/2023.06/eessi-2023.06-eb-4.9.2-2023a.yml

@@ -45,3 +45,5 @@ easyconfigs:
       options:
         # see https://github.com/easybuilders/easybuild-easyconfigs/pull/20880
         from-commit: bc6e08f89759b8b70166de5bfcb5056b9db8ec90
+  - libmad-0.15.1b-GCCcore-12.3.0.eb
+  - PyTorch-bundle-2.1.2-foss-2023a.eb

eb_hooks.py

@@ -285,6 +285,26 @@ def parse_hook_qt5_check_qtwebengine_disable(ec, eprefix):
         raise EasyBuildError("Qt5-specific hook triggered for non-Qt5 easyconfig?!")
 
 
+def parse_hook_sentencepiece(ec, eprefix):
+    """
+    LD_PRELOAD `libtcmalloc_minimal.so.4` on AARCH64-based systems
+    Avoids "libtcmalloc_minimal.so.4: cannot allocate memory in static TLS block" error
+    See https://github.com/EESSI/software-layer/pull/585/#issuecomment-2286068465
+    """
+    msg = f"Adding LD_PRELOAD to single sanity check command for {ec.name}"
+    print(msg)
+    if ec.name == "SentencePiece" and get_cpu_architecture() == AARCH64:
+        ec.log.info(msg)
+        # We need to tweak the sanitycheck command 'python -c 'import sentencepiece'" by
+        # adding the LD_PRELOAD. Since shell variables seem not get expanded we use a hardcoded path for now.
+        eessi_software_path = get_eessi_envvar('EESSI_SOFTWARE_PATH')
+        libtcmalloc_path = os.path.join(eessi_software_path, 'software', 'gperftools', '2.12-GCCcore-12.3.0', 'lib64', 'libtcmalloc_minimal.so')
+        scc = ec['sanity_check_commands']
+        scc_add_ld_preload = f"LD_PRELOAD={libtcmalloc_path} {scc[1]}"
+        ec['sanity_check_commands'][1] = scc_add_ld_preload
+    return ec
+
+
 def parse_hook_ucx_eprefix(ec, eprefix):
     """Make UCX aware of compatibility layer via additional configuration options."""
     if ec.name == 'UCX':
@@ -725,6 +745,34 @@ def post_sanitycheck_cuda(self, *args, **kwargs):
         raise EasyBuildError("CUDA-specific hook triggered for non-CUDA easyconfig?!")
 
 
+def post_sanitycheck_sentencepiece(self, *args, **kwargs):
+    """
+    LD_PRELOAD `libtcmalloc_minimal.so.4` on AARCH64-based systems
+    Avoids "libtcmalloc_minimal.so.4: cannot allocate memory in static TLS block" error
+    See https://github.com/EESSI/software-layer/pull/585/#issuecomment-2286068465
+    """
+    msg = f"Adding LD_PRELOAD to modluafooter for {self.name}"
+    print(msg)
+    if self.name == "SentencePiece" and get_cpu_architecture() == AARCH64:
+        # We want to set LD_PRELOAD so that it loads the libtcmalloc_minimal.so library from gperftools
+        # However, if LD_PRELOAD is already set, we need to prepend to it.
+        # An existing LD_PRELOAD can be space or colon separated, both are allowed
+        # So first we make sure it is colon seperated
+        # Colon seperation is allowed for LD_PRELOAD, so the easiest way to prepend to whats there
+        # is to specify it through modextrapaths
+        self.log.info(msg)
+        self.cfg['modluafooter'] = """
+libtcmalloc = pathJoin(os.getenv("EBROOTGPERFTOOLS"), "lib64", "libtcmalloc_minimal.so")
+prepend_path("LD_PRELOAD", libtcmalloc)
+"""
+
+
+def pre_sanitycheck_hook(self,*args, **kwargs):
+    """Main pre-sanitycheck hook: trigger custom functions based on software name."""
+    if self.name in PRE_SANITYCHECK_HOOKS:
+        PRE_SANITYCHECK_HOOKS[self.name](self, *args, **kwargs)
+
+
 def inject_gpu_property(ec):
     """
     Add 'gpu' property, via modluafooter easyconfig parameter
@@ -762,6 +810,7 @@ def inject_gpu_property(ec):
     'OpenBLAS': parse_hook_openblas_relax_lapack_tests_num_errors,
     'pybind11': parse_hook_pybind11_replace_catch2,
     'Qt5': parse_hook_qt5_check_qtwebengine_disable,
+    'SentencePiece': parse_hook_sentencepiece,
     'UCX': parse_hook_ucx_eprefix,
 }
 
@@ -804,6 +853,10 @@ def inject_gpu_property(ec):
     'numpy': post_single_extension_numpy,
 }
 
+PRE_SANITYCHECK_HOOKS = {
+}
+
 POST_SANITYCHECK_HOOKS = {
     'CUDA': post_sanitycheck_cuda,
+    'SentencePiece': post_sanitycheck_sentencepiece,
 }