Strange behavior trying to install in conda with PyTorch built from source

I'm running on a machine where I built PyTorch from source on CUDA 9.2 cuDNN 7.2.1, with Ubuntu 18.04. I built PyTorch from source exactly following directions, but when I went to install conda vision (conda install torchvision) I got this message:

jkterry@Station:~$ conda install torchvision
Solving environment: done

## Package Plan ##

  environment location: /home/jkterry/miniconda3

  added / updated specs: 
    - torchvision


The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    cudatoolkit-9.0            |       h13b8566_0       340.4 MB
    libtiff-4.0.9              |       he85c1e1_1         566 KB
    nccl-1.3.5                 |        cuda9.0_0         8.6 MB
    torchvision-0.2.1          |           py36_0          75 KB
    pillow-5.2.0               |   py36heded4f4_0         586 KB
    ninja-1.8.2                |   py36h6bb024c_1         1.3 MB
    olefile-0.46               |           py36_0          48 KB
    pytorch-0.4.1              |   py36ha74772b_0       215.8 MB
    cudnn-7.1.2                |        cuda9.0_0       367.8 MB
    ------------------------------------------------------------
                                           Total:       935.1 MB

The following NEW packages will be INSTALLED:

    cudatoolkit: 9.0-h13b8566_0      
    cudnn:       7.1.2-cuda9.0_0     
    libtiff:     4.0.9-he85c1e1_1    
    nccl:        1.3.5-cuda9.0_0     
    ninja:       1.8.2-py36h6bb024c_1
    olefile:     0.46-py36_0         
    pillow:      5.2.0-py36heded4f4_0
    pytorch:     0.4.1-py36ha74772b_0

The fact that torchvision is trying to install pytorch again seems to imply that the instructions to build from source don't properly integrate it into conda, or that the dependencies in torchvision are wrong. It appears to be torch vision, because of this:

jkterry@Station:~$ conda list
# packages in environment at /home/jkterry/miniconda3:
#
# Name                    Version                   Build  Channel
[...]
torch                     0.5.0a0+c8b246a           <pip>
[...]

Additionally, why is installing cuDNN and CUDA the default with conda install torchvision? Using torchivison doesn't require them, and accidentally adding a CUDA installation to a workstation can cause a huge number of issues.

Start by reproducing the shown conda install torchvision resolution in a matching environment and reviewing torchvision's package dependency metadata and source-build installation instructions. Done means determining whether the package incorrectly reinstalls PyTorch and CUDA/cuDNN, then documenting or correcting that behavior with a regression check if the repository provides one.