[BUG]: Conda installs a redundant CPU build of torch

[dagardner-nv]

Version

24.10

Which installation method(s) does this occur on?

Source

Describe the bug.

We install torch via pip which is how we get the 2.4.0+cu124 version.
I believe the sentence-transformers package is pulling in torchvision which in turn pulls in libtorch.

$ conda list | grep torch
libtorch                  2.4.0           cpu_generic_h4a3044c_1    conda-forge
torch                     2.4.0+cu124              pypi_0    pypi
torchdata                 0.8.0                    pypi_0    pypi
torchvision               0.19.1          cpu_py310hd9679db_0    conda-forge

Minimum reproducible example

CONDA_ALWAYS_YES=true conda env create --solver=libmamba -n morpheus -y --file conda/environments/all_cuda-125_arch-x86_64.yaml

Relevant log output

Click here to see error details

[Paste the error here, it will be hidden by default]

Full env printout

Click here to see environment details

[Paste the results of print_env.sh here, it will be hidden by default]

Other/Misc.

No response

Code of Conduct

• I agree to follow Morpheus' Code of Conduct
• I have searched the open bugs and have found no duplicates for this bug report

[efajardo-nv]

@dagardner-nv The sentence-transformers conda install (via examples_cuda-125_arch-x86_64.yaml) in the release container results in the following pytorch packages in the container:

# packages in environment at /opt/conda/envs/morpheus:
#
# Name                    Version                   Build  Channel
libtorch                  2.4.1           cpu_generic_hb3b73e9_0    conda-forge
pytorch                   2.4.1           cpu_generic_py310hcbfaffa_0    conda-forge
torch                     2.4.0+cu124              pypi_0    pypi
torchvision               0.19.1          cpu_py310h0339c84_1    conda-forge

The VDB embedding stage then chooses to use CPU version. Replacing with pip package switches it back to GPU but it's not quite as fast (~3 min vs ~2 min in our example).