diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6b80b24..563122a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -83,7 +83,7 @@ foreach(i ${rochpl_device_source})
 endforeach()
 
 # HIP flags workaround while target_compile_options does not work
-list(APPEND HIP_HIPCC_FLAGS "-Wno-unused-command-line-argument -Wno-deprecated-declarations -fPIE -fopenmp")
+list(APPEND HIP_HIPCC_FLAGS "-Wno-unused-command-line-argument -Wno-deprecated-declarations -fPIE -fopenmp --gpu-max-threads-per-block=1024")
 list(APPEND CMAKE_HOST_FLAGS "-Wno-deprecated-declarations")
 
 if (CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -94,25 +94,51 @@ else()
   list(APPEND CMAKE_HOST_FLAGS "-O3;-march=native")
 endif()
 
-# GPU arch targets
-set(TARGETS "gfx900;gfx906")
-if(HIP_VERSION VERSION_GREATER_EQUAL "3.7")
-  set(TARGETS "${TARGETS};gfx908")
-endif()
-if(HIP_VERSION VERSION_GREATER_EQUAL "4.3")
-  set(TARGETS "${TARGETS};gfx90a")
-endif()
-if (HIP_VERSION VERSION_GREATER_EQUAL "5.7")
-  set(TARGETS "${TARGETS};gfx942")
-endif()
-if (HIP_VERSION VERSION_GREATER_EQUAL "6.5")
-  set(TARGETS "${TARGETS};gfx950;gfx1100")
+set(ARCHS "")  # use plural to indicate list
+if(DEFINED HPL_BUILD_ARCH AND NOT HPL_BUILD_ARCH STREQUAL "")
+  string(REPLACE "," ";" ARCHS "${HPL_BUILD_ARCH}")
+  list(TRANSFORM ARCHS STRIP)
+  list(REMOVE_DUPLICATES ARCHS)
+  message(STATUS "Using manually specified GPU targets: ${ARCHS}")
+else()
+  message(STATUS "Detecting available architecture")
+  ############ Find using rocminfo #####################
+  find_program(ROCMINFO_EXECUTABLE rocminfo)
+  if(ROCMINFO_EXECUTABLE)
+    execute_process(
+      COMMAND ${ROCMINFO_EXECUTABLE}
+      OUTPUT_VARIABLE ROCMINFO_OUTPUT
+      ERROR_QUIET
+      OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+
+    # 1) Only match lines where the token follows "Name:"
+    string(REGEX MATCHALL "Name:[ \t]+gfx[0-9a-z]+" ARCH_MATCHES "${ROCMINFO_OUTPUT}")
+
+    # 2) Strip the leading "Name:   " to keep just gfx tokens
+    string(REGEX REPLACE "Name:[ \t]+" "" ARCHS "${ARCH_MATCHES}")
+
+    # 3) Remove duplicates
+    list(REMOVE_DUPLICATES ARCHS)
+
+    foreach(match ${ARCHS})
+      string(REGEX REPLACE "Name:\\s+" "" arch "${match}")
+      list(APPEND ARCH "${arch}")
+    endforeach()
+  endif()
 endif()
 if (HIP_VERSION VERSION_GREATER_EQUAL "7.0")
   set(TARGETS "${TARGETS};gfx1201")
 endif()
 
-foreach(target ${TARGETS})
+if(ARCHS STREQUAL "")
+    message(FATAL_ERROR "No GPU architectures detected via rocminfo and no BUILD_ARCH specified. Use ./install.sh --arch=gfxXXX")
+endif()
+
+message(STATUS "Building for GPU architecture: ${ARCHS}")
+
+# Generate HIP_HIPCC_FLAGS
+foreach(target ${ARCHS})
   list(APPEND HIP_HIPCC_FLAGS "--offload-arch=${target}")
 endforeach()
 
diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
index 6d6be5d..468420e 100644
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -101,7 +101,8 @@ if(NOT ROCM_FOUND)
   execute_process(COMMAND ${CMAKE_COMMAND} -E tar xzf ${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag}.zip
                   WORKING_DIRECTORY ${PROJECT_EXTERN_DIR})
 
-  find_package(ROCmCMakeBuildTools REQUIRED CONFIG PATHS ${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag})
+  # find_package(ROCmCMakeBuildTools REQUIRED CONFIG PATHS ${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag})
+  set(CMAKE_MODULE_PATH "${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag}/share/rocm/cmake;${CMAKE_MODULE_PATH}")
 endif()
 
 include(ROCMSetupVersion)
diff --git a/install.sh b/install.sh
index b30a3fb..9ada030 100755
--- a/install.sh
+++ b/install.sh
@@ -17,6 +17,7 @@ function display_help()
   echo "    [--with-rocm=<dir>] Path to ROCm install (Default: /opt/rocm)"
   echo "    [--with-rocblas=<dir>] Path to rocBLAS library (Default: /opt/rocm/rocblas)"
   echo "    [--with-mpi=<dir>] Path to external MPI install (Default: clone+build OpenMPI)"
+  echo "    [--arch] Specify comma separated architecture list to build (Default: detect from rocm_agent_enumerator)"
   echo "    [--with-mpi-gtl=<dir>] Path to external MPI-GTL install (Optional: defaults to no gtl support)"
   echo "    [--verbose-print] Verbose output during HPL setup (Default: true)"
   echo "    [--progress-report] Print progress report to terminal during HPL run (Default: true)"
@@ -33,7 +34,7 @@ supported_distro( )
   fi
 
   case "${ID}" in
-    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos)
+    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos|kylin|rocky)
         true
         ;;
     *)  printf "This script is currently supported on Debian, Linuxmint, Ubuntu, CentOS, RHEL, Fedora and SLES\n"
@@ -68,11 +69,11 @@ exit_with_error( )
         printf "sudo apt install -y ${library_dependencies_ubuntu[*]}\n"
         ;;
 
-      centos|rhel|tencentos)
+      centos|rhel|tencentos|kylin)
         printf "sudo yum -y --nogpgcheck install ${library_dependencies_centos[*]}\n"
         ;;
 
-      fedora)
+      fedora|rocky)
         printf "sudo dnf install -y ${library_dependencies_fedora[*]}\n"
         ;;
 
@@ -152,6 +153,11 @@ install_openmpi( )
     exit 3
   fi
 
+  UCX_ROOT="$(pwd)/tpl/ucx"
+  export LD_LIBRARY_PATH="${UCX_ROOT}/lib:${UCX_ROOT}/lib64:${LD_LIBRARY_PATH}"
+  export LIBRARY_PATH="${UCX_ROOT}/lib:${UCX_ROOT}/lib64:${LIBRARY_PATH}"
+  export CPATH="${UCX_ROOT}/include:${CPATH}"
+
   if [ ! -d "./tpl/openmpi" ]; then
     mkdir -p tpl && cd tpl
     git clone --branch v5.0.7 --recursive https://github.com/open-mpi/ompi.git openmpi
@@ -232,7 +238,7 @@ enable_tracing=false
 # check if we have a modern version of getopt that can handle whitespace and long parameters
 getopt -T
 if [[ $? -eq 4 ]]; then
-  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,debug,prefix:,with-rocm:,with-mpi:,with-mpi-gtl:,with-rocblas:,verbose-print:,progress-report:,detailed-timing:,enable-tracing: --options hg -- "$@")
+  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,debug,prefix:,with-rocm:,with-mpi:,with-mpi-gtl:,with-rocblas:,verbose-print:,arch:,progress-report:,detailed-timing:,enable-tracing: --options hg -- "$@")
 else
   echo "Need a new version of getopt"
   exit_with_error 1
@@ -263,6 +269,9 @@ while true; do
     --with-mpi)
         with_mpi=${2}
         shift 2 ;;
+    --arch)
+    arch=${2}
+    shift 2 ;;
     --with-mpi-gtl)
         with_mpi_gtl=${2}
         shift 2 ;;
@@ -347,6 +356,9 @@ pushd .
   if [[ "${enable_tracing}" == on || "${enable_tracing}" == true || "${enable_tracing}" == 1 || "${enable_tracing}" == enabled ]]; then
     cmake_common_options="${cmake_common_options} -DHPL_TRACING=ON"
   fi
+  if [[ -n "${arch}" ]]; then
+    cmake_common_options="${cmake_common_options} -DHPL_BUILD_ARCH=${arch}"
+  fi
   shopt -u nocasematch
 
   # Build library with AMD toolchain because of existence of device kernels
diff --git a/scripts/mpirun_rochpl.in b/scripts/mpirun_rochpl.in
index 155f502..05a96e0 100755
--- a/scripts/mpirun_rochpl.in
+++ b/scripts/mpirun_rochpl.in
@@ -46,7 +46,7 @@ supported_distro( )
   fi
 
   case "${ID}" in
-    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos)
+    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos|kylin|rocky)
         true
         ;;
     *)  printf "This script is currently supported on Debian, Linuxmint, Ubuntu, CentOS, RHEL, Fedora and SLES\n"
@@ -101,6 +101,13 @@ filename=HPL.dat
 inputfile=false
 cmdrun=false
 
+tpl_dir=$(dirname "$(readlink -f "$0")")/../tpl
+ompi_prefix=$tpl_dir/openmpi
+ompi_lib_dir=$tpl_dir/openmpi/lib
+ucx_lib_dir=$tpl_dir/ucx/lib
+
+export LD_LIBRARY_PATH=$ompi_lib_dir:$ucx_lib_dir:$LD_LIBRARY_PATH
+export OPAL_PREFIX=$ompi_prefix
 devices=
 
 # #################################################
@@ -117,12 +124,16 @@ mpi_args=
 #Check if using OpenMPI
 if [[ $(${mpi_bin} --version | grep "open-mpi") ]]; then
   mpi_args+=" --map-by node --rank-by slot --bind-to none "
+  #mpi_args+=" --map-by numa:PE=16 --bind-to core --report-bindings "
 
   #Check if this is OpenMPI+UCX
   ompi_info=$(dirname ${mpi_bin})/ompi_info
   if [[ $(${ompi_info} | grep "MCA pml: ucx") ]]; then
     # ucx-specific args
-    mpi_args="--mca pml ucx --mca btl ^vader,tcp,openib,uct ${mpi_args}"
+    mpi_args="--mca pml ucx --mca btl ^vader,tcp,openib,uct \
+              -x UCX_TLS=self,sm,rocm_ipc,rocm_copy,rc_mlx5 \
+              -x UCX_MEMTYPE_CACHE=n \
+              ${mpi_args}"
   fi
 fi
 
@@ -153,7 +164,7 @@ while true; do
         exit 0
         ;;
     --version)
-        ${mpi_bin} -np 1 ${mpi_args} ${rochpl_runscript} --version
+        ${mpi_bin} --allow-run-as-root -np 1 ${mpi_args} ${rochpl_runscript} --version
         exit 0
         ;;
     -P)
@@ -219,4 +230,4 @@ if [ ! -z "${devices}" ]; then
 fi
 
 #run
-${mpi_bin} -np ${np} ${mpi_args} ${rochpl_runscript} ${rochpl_args}
+${mpi_bin} --allow-run-as-root -np ${np} ${mpi_args} ${rochpl_runscript} ${rochpl_args}
diff --git a/scripts/run_rochpl.in b/scripts/run_rochpl.in
index 1522e5d..1380d3a 100755
--- a/scripts/run_rochpl.in
+++ b/scripts/run_rochpl.in
@@ -46,7 +46,7 @@ supported_distro( )
   fi
 
   case "${ID}" in
-    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos)
+    debian|linuxmint|ubuntu|centos|rhel|fedora|sles|tencentos|kylin|rocky)
         true
         ;;
     *)  printf "This script is currently supported on Debian, Linuxmint, Ubuntu, CentOS, RHEL, Fedora and SLES\n"
@@ -104,7 +104,9 @@ cmdrun=false
 
 devices=
 
-export LD_LIBRARY_PATH=${rocblas_dir}:${blas_dir}:${rocm_dir}/lib:$LD_LIBRARY_PATH
+tpl_dir=$(dirname "$(readlink -f "$0")")/../tpl
+ucx_lib_dir=$tpl_dir/ucx/lib
+export LD_LIBRARY_PATH=${rocblas_dir}:${blas_dir}:$ucx_lib_dir:${rocm_dir}/lib:$LD_LIBRARY_PATH
 
 # #################################################
 # Parameter parsing
@@ -274,7 +276,7 @@ myq=$((rank/p))
 cpulist=$(lscpu --parse=CPU,CORE,NODE | awk '!/#/' | tr ',' "\t" | sort -k 2 -g -s)
 
 #construct list of devices and their numa affinities
-devicelist=$(${rocm_dir}/bin/rocm-smi --csv --showtoponuma | tail -n +2 | tr ',' "\t")
+devicelist=$(hy-smi --csv --showtoponuma | tail -n +2 | tr ',' "\t")
 
 #count the cpus per core
 threads_per_core=$(echo "${cpulist}" | grep -c ".*	0	.*")
@@ -363,7 +365,8 @@ export OMP_PROC_BIND=true
 
 
 if [[ $globalRank -lt $size ]]; then
-  echo "Node Binding: Process $rank [(p,q)=($myp,$myq)] GPU: $mygpu, CPU Cores: $omp_num_threads - $places"
+  echo "Node Binding: Process $rank [(p,q)=($myp,$myq)] GPU: $mygpu, NUMA: $mynuma, CPU Cores: $omp_num_threads - $places"
+#  echo "Node Binding: Process $rank [(p,q)=($myp,$myq)] GPU: $mygpu, NUMA: $mynuma"
 fi
 
 rochpl_args="-P ${P} -Q ${Q} -p ${p} -q ${q} -f ${frac} -it ${it}"
@@ -374,4 +377,5 @@ else
 fi
 
 #run
-${rochpl_bin} ${rochpl_args}
+#${rochpl_bin} ${rochpl_args}
+numactl --cpunodebind=${mynuma} --membind=${mynuma} ${rochpl_bin} ${rochpl_args}
diff --git a/src/HPL_pdtest.cpp b/src/HPL_pdtest.cpp
index 94a0d3f..3135763 100644
--- a/src/HPL_pdtest.cpp
+++ b/src/HPL_pdtest.cpp
@@ -212,7 +212,7 @@ void HPL_pdtest(HPL_T_test* TEST,
                     ctime(&current_time_end));
       }
 #ifdef HPL_PROGRESS_REPORT
-      printf("Final Score:    %7.4e GFLOPS \n", Gflops);
+      printf("Final Score:    %7.9e GFLOPS \n", Gflops);
 #endif
     }
 #ifdef HPL_DETAILED_TIMING
diff --git a/src/pgesv/HPL_pdgesv.cpp b/src/pgesv/HPL_pdgesv.cpp
index d6c99c3..280a9a5 100644
--- a/src/pgesv/HPL_pdgesv.cpp
+++ b/src/pgesv/HPL_pdgesv.cpp
@@ -336,7 +336,7 @@ void HPL_pdgesv(HPL_T_grid* GRID, HPL_T_palg* ALGO, HPL_T_pmat* A) {
       printf("  %9.3e  |", step_gflops);
 #endif
 
-      printf("    %9.3e   \n", gflops);
+      printf("    %9.9e   \n", gflops);
     }
 #endif