Open Issues Need Help
View All on GitHubAI Summary: This task aims to automate the assignment of the 'organization' tag for training runs, which is crucial for cost analysis between different themes. It involves mapping a developer's HPC username and HPC name to their organization, using a new, private configuration file separate from existing private configs. The implementation should be within the platform env module of the WeatherGenerator project and support specific Tier 1 clusters.
AI Summary: This GitHub issue proposes enhancing the `clean.py` script to include two new cleaning targets: `uv_cache_dir` contents older than 48 hours and `slurm_dir` contents older than 3 months. The update requires implementing a dry-run option, specific command-line flags to control cleaning for each target, and a careful testing protocol before deployment to prevent data loss.
AI Summary: The task is to explicitly cap the number of training nodes used by the `launch-slurm` script to a maximum of 2, as there is currently no performance benefit beyond this limit. This involves adding a `--nodes` option to the script; if a value greater than 2 is provided, it should trigger a hard error, otherwise, the specified node count will be passed to Slurm.
AI Summary: During model checkpoint uploads using `upload_experiment.py`, the terminal repeatedly prints `InsecureRequestWarning` from `urllib3` (O(100) times), even though the uploads complete successfully. The user requests either suppressing these warnings or implementing proper certificate verification to resolve the issue.
AI Summary: Migrate the `config` module from `weathergen.utils` to `weathergen.common` to improve dependency management. This involves updating import paths in all modules currently using `utils.config`.
AI Summary: The task is to improve the naming convention for data streams within the WeatherGenerator project to ensure compatibility with logging systems like MLflow and WandB. The current system uses descriptive names that are not suitable for these systems. The solution involves choosing between converting descriptive names to a compatible format, enforcing a stricter naming convention, or requiring users to provide a separate machine-readable identifier for each stream.