February 2022 - Ecp-aurora-sdk-discuss

New Frameworks release is available
by Ramer, Rebecca 11 Feb '22

11 Feb '22

We would like to announce availability of a new frameworks module for pytorch and tensorflow for use in the JLSE. This release is based on the 2022.1.0 NDA version of the AiKit. Information on how to setup and work with this release can be found here for pytorch: https://wiki.jlse.anl.gov/display/inteldga/AIKit+NDA+-+2022.1.0+-+Pytorch and here for tensorflow: https://wiki.jlse.anl.gov/display/inteldga/AIKit+NDA+-+2022.1.0+-+Tensorflow At this time, the pytorch adorym model is not functional, an active investigation is underway. More details about model runs, and running unit tests will be available on the wiki pages next week. Also, further effort is needed to enable the distributed training model runs. We hope to enable this functionality soon. Please let us know of any issues encountered. Rebecca Ramer rebecca.ramer(a)intel.com<mailto:[email protected]>

1 0

New version of mpich on JLSE
by Chan-nui, Christopher 11 Feb '22

11 Feb '22

Hello, We are pleased to announce that an update to MPICH drop 45.3 is available by default on JLSE. If you wish to force loading MPICH without GPU support you may do so by using the following command: ONEAPI_MPICH_GPU=NO_GPU module add oneapi Release 3 has the following changes: * Remove linking of libze_loader.so from nogpu RPMs * Remove linking of libimf.so and libze_loader.so from hydra_pmi_proxy * Fix an issue with pmix RPMs not being relocatable -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

New KMD update to Arcticus nodes
by Muralidharan, Servesh 11 Feb '22

11 Feb '22

Hello Arcticus testbed users, A new KMD update from agama-prerelease-335 is being rolled out across the arcticus nodes. The update is expected to complete by end of today. No user specific changes are needed, updates will progress on a rolling fashion. Once the process is complete nodes will be marked back up. On behalf of early hardware team. -- Servesh

1 0

Updated kokkos/openmptarget_inel and kokkos/sycl_intel modules to the latest oneapi (2022.01.30.002 eng-compiler)
by Williams, Timothy J. 01 Feb '22

01 Feb '22

For those using the centrally-installed kokkos modules in /soft/restricted/CNDA/modulefiles: Using the new SDK in the new default oneapi module, I built new kokkos versions for {Serial,OpenMP,OpenMPTarget} execution spaces (“backends”) and {Serial,OpenMP,SYCL} execution spaces, for JIT and AoT. These are what you get with the usual default modules: module load kokkos/openmptarget_intel module load kokkos/openmptarget_intel_aot module load kokkos/sycl_intel module load kokkos/sycl_intel_aot If you are not using the new default oneapi module, you will need to explicitly load earlier versions of the kokkos modules to match your oneapi version. The prior default kokkos versions were: module load kokkos/2021.10.30.003/eng-compiler/openmptarget_intel module load kokkos/2021.10.30.003/eng-compiler/openmptarget_intel _aot module load kokkos/2021.10.30.003/eng-compiler/sycl_intel module load kokkos/2021.10.30.003/eng-compiler/sycl_intel_aot Best regards, Tim ------- Timothy J. Williams, Ph.D. --- Argonne National Laboratory --- tjwilliams(a)anl.gov -------

1 0