feat: parallel inference without slurm

[cathalobrien]

This PR extends the Parallel Inference added in #55 to work without slurm within 1 node. It's much nicer to debug and run now :D

When running parallel inference without slurm, you have to add world_size to the config. At the moment, this is ignored in favour of SLURM_NTASKS when running with srun. An example of a config running parallel inference across 4 nodes is shown below. You can launch this job as normal with anemoi-inference run parinf.yaml.

checkpoint: /path/to/inference-last.ckpt
lead_time: 60
runner: parallel
world_size: 4 #Only required if running parallel inference without Slurm
input:
  grib: /path/to/input.grib
output:
  grib: /path/to/output.grib

How it works is:

• check if anemoi-inference is launched by srun
• if not, spawn config.world_size processes
  • master_addr is localhost
  • master_port is a hash of the node name, within a range.
• Each spawned process runs a slimmed down version RunCmd.run but with the config preloaded

Issues:

• The master port calculation would lead to a clash if two parallel inference processes ran on the same node at the same time.
• At the moment, I have a copy of RunCmd.run in runners/parallel.py. Would be nice to be able to use that code directly, rather then having to maintain a copy. To do this, I would just have to be able to pass a loaded config instead of a path

---

📚 Documentation preview 📚: https://anemoi-inference--121.org.readthedocs.build/en/121/

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.03%. Comparing base (90728d5) to head (9d22d57).
Report is 44 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #121   +/-   ##
=======================================
  Coverage   98.03%   98.03%           
=======================================
  Files           3        3           
  Lines          51       51           
=======================================
  Hits           50       50           
  Misses          1        1           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.