Support

Start from the route that matches the problem. pimm currently provides public GitHub issue tracking; the repository does not document a chat channel, response-time guarantee, or paid support program.

Need	First action
install, data, launch, training, checkpoint, or evaluation failure	search Troubleshooting, then search GitHub issues
reproducible software bug	open a GitHub issue with the minimal reproducer below
documentation error	open an issue and link the page/heading plus the code or command that disagrees
feature or research-method proposal	open an issue describing the user outcome, scientific contract, compatibility impact, and bounded validation plan
private data/access/account problem	contact the owner of that dataset or computing site; do not post credentials or private paths publicly
question about a published model/result	include the exact Hub repository/revision, config, dataset revision, and metric definition

Before opening an issue

git rev-parse HEAD
git status --short
uv lock --check
uv run pimm launch --train.config <config> --dry-run

Search the exact first error, not only the final launcher exit. Reduce the run to one process, a bounded event subset, and the smallest config change that still fails when that is scientifically safe.

Minimal bug report

Copy this checklist into the issue:

### Expected behavior
[What should have happened?]

### Actual behavior
[First error and relevant preceding log lines]

### Reproducer
[Exact command and smallest config/data shape summary]

### Environment
- pimm version and full Git commit:
- modified worktree: yes/no (attach relevant diff)
- OS, Python, PyTorch, CUDA runtime, NVIDIA driver:
- GPU model/count; world size and workers:
- install path: locked source / standard image / NERSC image / other:

### Inputs
- config or resolved_config.json:
- checkpoint URI + immutable revision/checksum:
- dataset type, revision, split, and minimal schema/shapes:

### Distributed context (if applicable)
- rank/node/job ID of first failure:
- scheduler output and rendezvous/NCCL error:

Redact tokens, credentials, private hostnames, usernames, storage paths, and unreleased data. Do not attach a full dataset when a schema, synthetic fixture, or a few authorized events reproduce the failure.

Scientific questions need provenance

“My metric differs” is not actionable without the checkpoint revision, resolved preprocessing, dataset revision/split, class map, ignored labels, metric aggregation, and evaluation code revision. Compare those before tuning the model. The checklist in Evaluate is the minimum useful context.

Security and private reports

The repository does not currently include SECURITY.md or a documented private vulnerability-reporting address. Do not put credentials, exploitable details involving private infrastructure, or sensitive data in a public issue.

TODO

Publish a security policy with supported versions and a private reporting route. Until then, contributors must obtain an appropriate private contact from the DeepLearnPhysics/project maintainers before sharing sensitive details.

Contribution path

If the fix is understood and scoped, follow Contributing. Opening a pull request does not require a preceding issue for a small documentation or test fix, but larger behavior/scientific changes benefit from agreement on the contract first.