inference quickstart

[mahf708]

this is a quick-start to inference, discussed yesterday... for now only a draft to get feedback, then we will add some more info (especially about formalizing a process for "simulation pages" on confluence). To reviewing resulting doc, see this page https://docs.e3sm.org/aigroup/pr-preview/pr-19/ace2-inference/

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://E3SM-Project.github.io/aigroup/pr-preview/pr-19/
Built to branch gh-pages at 2026-01-30 20:54 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[wlin7]

Thanks @mahf708 for putting up this inference quickstart. I faithfully followed the steps and obtained a successful test1.

After going through the process, I feel it will be useful to have some additional guides:

1. How to create an environment that contains pytorch and uv, and ensuring them running with the same python interpreter. What other packages are suggested to also be included. I likely started with an unclean environment, causing uv to run a different python interpreter. I had to force the UV_PYTHON to get it to work.

2. Need to request a gpu node to run the uv command, at least for perlmutter. Though the login node also has a GPU, running on a login node would get CUDA out of memory error. Should n_initial_conditions always equal to num_data_workers? I requested 2 gpus on 1 node to run the example. Are they a suitable match (best use of the requested resource)?

3. restart.nc file is recorded. How to perform a restart run?

4. Some info about the output data: what are the pressure values for levels 0 to 7? How the monthly_mean data are computed? Equal length for all months? (The example test1 ran 1000 steps -- 250 days, and has 11 monthly values)

[wlin7]

pyTorch as well.

[mahf708]

@wlin7 I added some info about python envs separately; the expectation is that people will figure this out, but uv is used as a shortcut. In the future, we can potentially host an aigroup env like e3sm-unified so that people can just activate it

[mahf708]

Hi @elynnwu, do you know what the restart.nc does in the context of ACE? I didn't look into it yet. Also, adding this quickstart for people to try out inference using ACE, any comments appreciated!

[elynnwu]

do you know what the restart.nc does in the context of ACE?

It's the last prognostic state in inference, so it can be used as a new initial condition file if you want to run another inference from this step.