Getting help#

FlashDreams is an open-source project. Picking the right channel up-front gets you a useful answer fastest.

Choose a channel#

I think I found a bug

File a GitHub issue with the smallest reproducer you can manage. See the checklist below for what makes a bug report easy to act on.

https://github.com/NVIDIA/flashdreams/issues
I have a question about how to do X

Discord is the venue for open-ended “how do I…” questions, sharing results, and office hours with maintainers and other users. The issue tracker with the question label is fine too.

Discord
I have a feature idea

Open an issue describing the use case, what you’d want the API to look like, and the trade-offs you can think of. For larger features, please discuss before sending a PR.

https://github.com/NVIDIA/flashdreams/issues/new
I found a security issue

Do not file as a public issue. Follow the coordinated disclosure process in SECURITY.md.

https://github.com/NVIDIA/flashdreams/blob/main/SECURITY.md

Before you file an issue#

  • Search existing issues (open and closed) for your error message or symptom. Most “is this a bug?” questions already have an answer.

  • Check the FAQ page. If your question is there, great; if a related question is there, link to it in your issue.

  • Check the Troubleshooting page. It lists common first-run failures (e.g. CUDA build mismatches, disk and cache limits, Hugging Face auth, GPU memory) with a likely cause and next step for each.

  • Confirm your version. A bug fixed in main looks identical to a fresh bug if you’re on an older tagged release. Reproduce against the latest main or note your version in the report.

  • Try with the smallest possible inputs. A 5-minute repro on a single GPU is more actionable than a multi-node training job.

What makes a good bug report#

  • What you ran. The exact command, recipe name, or Python snippet.

  • What you expected. One sentence.

  • What you saw. Full stack trace or output. Wrap it in a code fence; don’t paste a screenshot of text.

  • Your environment. Python version, CUDA version, GPU model, FlashDreams version (python -c "import flashdreams; print(flashdreams.__version__)"), and how you installed it (workspace checkout, pip install, container image).

  • What you’ve already tried. Workarounds, related issues, debug prints — any of these speed up triage.

Response times#

The maintainers aim for a first review on every PR within two business days (see the Contributing to FlashDreams guide for the canonical statement). Issues have no formal service-level agreement.

Commercial support#

FlashDreams is offered as-is under the Apache-2.0 license. There is no commercial support agreement attached to the open-source project.