I've noticed that mlflow is super slow at times to install using the existing environment, and einops and transformers are really slow as well. I hadn't figured out why but removing these made installing everything else reasonable. I had issues on Windows that the environment would also default to CPU Pytorch for some reason. On the Pytorch site the conda command now includes pytorch-cuda as a package with a specification for a version of CUDA so that possibly should be added here as well.

A bare-minimum environment that appears to work is:

name: monai
channels:
  - pytorch
  - defaults
  - nvidia
dependencies:
  - numpy>=1.17
  - pytorch>=1.6
  - pytorch-cuda=11.6

I also created a runtime version of the environment with every package for development removed, this was a little slimmer for just deploying MONAI somewhere, but its contents will change now with the new way Pytorch is doing its packages.

I just tested this works fine:

name: monai
channels:
  - pytorch
  - defaults
  - conda-forge
dependencies:
  - numpy>=1.17
  - pytorch>=1.8
  - pip
  - pip:
    - -r requirements-dev.txt

https://github.com/Project-MONAI/MONAI/actions/runs/3564985942

assuming pytorch is installed following the 'conda' option here https://pytorch.org/get-started/locally/#start-locally

In the logs I see that the Windows build chose the cpuonly package by default. This is specified in conda.yml but I've seen this happening as a default thing when installing in Windows and I've always had to setup environments manually there.

In the logs I see that the Windows build chose the cpuonly package by default. This is specified in conda.yml but I've seen this happening as a default thing when installing in Windows and I've always had to setup environments manually there.

I think manually installing pytorch is the recommended way described here https://pytorch.org/get-started/locally/#start-locally let me try to include the cuda requirement for the ci..

I've added the cuda dependency following your minimum environment example @ericspod, it works fine with ubuntu and windows: https://github.com/Project-MONAI/MONAI/actions/runs/3566365419/jobs/5992720823
but is not compatible with macos, so I drop the macOS test.

Apart from the downloading took some time, it looks good to me. Any further suggestions? @ericspod @rijobro. Dropping these pip package requirements in the yaml will make it faster,

  - pip:
    - -r requirements-dev.txt

I can do that, just wanted to get your opinion as it'll be very different from the current behaviour of conda update -f environment-dev.yml.

I've added the cuda dependency following your minimum environment example @ericspod, it works fine with ubuntu and windows: https://github.com/Project-MONAI/MONAI/actions/runs/3566365419/jobs/5992720823 but is not compatible with macos, so I drop the macOS test.

Apart from the downloading took some time, it looks good to me. Any further suggestions? @ericspod @rijobro. Dropping these pip package requirements in the yaml will make it faster,

I think this looks good but I want to test this at home with my home Windows box I know I've had issues with. I was wondering if we wanted to do multiple versions of the environment file to have everything, only packages needed for deploy, and a minimal package.

I've run with the environment file on my Windows box and it's installed things as I expected with the GPU version of Pytorch.

/build