* Added hash values for LBANN v0.101 and Hydrogen v1.5.0. Updated the
LBANN package to be more successful in resolving a legal configuration
of MPI and HWLOC packages. This required the removal of the MPI
virtual package since it is unable to resolve dependencies with
minimum version requirements. As a result to enable a reasonable
install line for LBANN this requires explicit forwarding of MPI
variants to Hydrogen and Aluminum. Due to the lack of variant
forwarding, there are many explicitly replicated dependencies for both
LBANN and Hydrogen. Fixed the error in LBANN where gpu variant was
replaced by the cuda variant, but not all dependencies were fixed.
* Fixed the minumum cuDNN version for newer versions of LBANN.
* Added explicit versioning of the MPI libraries for DiHydrogen to avoid
all of the conflicts with minimum required versions of the OpenMPI library.
* Removed explicit MPI versions and went back to using the MPI virtual
dependency. Updated construction of variant forwarding to use
iterative construction of constraints and variants. This exacerbates
the challenges with backtracking in the current concretizer, but
should be fixed in the new concretizer.
* Added support for including the DiHydrogen library in LBANN as well as
support for the distributed convolution (DistConv) parallel
algorithms. Also include support for building with half precision.
* Moving dependencies around
* Added conflict statement to ensure that the variant dihydrogen is
required for distconv.
* Removed the preferred field
* Fixed Flake8 and cuDNN version bounds
* Update LBANN, Hydrogen, Aluminum to inherit CudaPackage
* Update CMake constraints: LBANN, Hydrogen, and Aluminum now require
cmake@3.16.0: (better support for pthreads with nvcc)
* Aluminum: add variants for host-enabled MPI and RMA features in a
MPI-GPU RDMA-enabled library
* NCCL: add versions 2.7.5-1, 2.7.6-1, and 2.7.8-1
* Hydrogen: add version 1.4.0
* LBANN: add versions 0.99 and 0.100
* Aluminum: add versions 0.4.0 and 0.5.0
* Added a package for the DiHydrogen distributed linear algebra library.
* Updated recipe to provide cuda architecture constaints.
* Addressed reviewer comments
* Fixed flake 8
