Handling of PyTorch dependency

[jakirkham]

During a recent review of dependencies here we noticed that PyTorch is a dependency, but is not listed as one at runtime

wholegraph/conda/recipes/pylibwholegraph/meta.yaml

Lines 68 to 74 in 9b877a4

	run:
	- {{ pin_compatible('cuda-version', max_pin='x', min_pin='x') }}
	{% if cuda_major == "11" %}
	- cudatoolkit
	{% endif %}
	- libwholegraph ={{ version }}
	- python

However there are a range of tightly pinned PyTorch builds (on the older side). All of them for CUDA 11. No CUDA 12

Details

wholegraph/dependencies.yaml

Lines 230 to 291 in 9b877a4

	specific:
	- output_types: [conda]
	matrices:
	- matrix:
	arch: x86_64
	cuda: "11.2"
	packages:
	# It's impossible to create this environment with pyg because
	# the pyg package has an explicit dependency on cudatoolkit=11.*
	# and there simply isn't any build for cudatoolkit=11.2.
	# Note that the packages for CUDA 11.2/11.4 environments are the
	# ones from conda-forge (built only against CUDA 11.2) and
	# *not* the pytorch channel. For CUDA 11.5/11.8 environments,
	# we're using packages from the pytorch channel.
	- pytorch=1.11.0=*cuda112*
	- matrix:
	arch: x86_64
	cuda: "11.4"
	packages:
	# It's impossible to create this environment with pyg because
	# the pyg package has an explicit dependency on cudatoolkit=11.*
	# and there simply isn't any build for cudatoolkit=11.4.
	# There is also no build of pytorch for CUDA 11.4 but the 11.2
	# build should work in practice and doesn't require any
	# cudatoolkit version explicitly.
	- pytorch=1.11.0=*cuda112*
	- matrix:
	arch: x86_64
	cuda: "11.5"
	packages:
	# This environment "just works" for both pytorch and pyg, but only
	# with older pytorch versions since the newest ones aren't built
	# against 11.5 anymore.
	- pytorch=1.11.0=*cuda11.5*
	- matrix:
	arch: x86_64
	cuda: "11.8"
	packages:
	# Since CUDA 11.6, pytorch switched to using the `cuda-*` packages
	# as dependencies for its official conda package. These are only
	# available from the nvidia channel at the moment, and this will
	# probably continue once conda-forge has added these new packages
	# since conda-forge will only add this from CUDA 12.0 onwards,
	# at least in the near-term.
	# Our own RAPIDS packages are dependent on the `cudatoolkit`
	# package from conda-forge though, which means that we have to
	# install both `cudatoolkit` version 11.8 and the `cuda-*` packages
	# version 11.8 here.
	# Starting with Pytorch 2.0, this works well though, since Pytorch
	# has largely reduced its dependencies, so only part of the CUDA
	# toolkit needs to be duplicated this way.
	# If conda-forge supports the new cuda-* packages for CUDA 11.8
	# at some point, then we can fully support/properly specify
	# this environment.
	- pytorch=2.0.0
	- pytorch-cuda=11.8
	- matrix:
	arch: aarch64
	cuda: "11.8"
	packages:
	- pytorch=2.0.0
	- pytorch-cuda=11.8

Not really sure how to handle updating PyTorch here or how it should be reflected in dependencies. So opening this issue to discuss and find a reasonable resolution