Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Contributing

This is a community wiki for the gfx1030 (RDNA2 / Navi 21) GPU ecosystem. Contributions of all sizes are welcome — fixing a typo, correcting an out-of-date command, or adding a whole new guide.

Live discussion happens on the gfx1030 club Discord: https://discord.gg/mESex2aBp. Propose wiki fixes there (#wiki) or open a GitHub PR — both are fine. See Useful resources.

A daily Cursor Automation (when enabled) reads Discord and opens draft PRs titled wiki: discord digest YYYY-MM-DD. Treat those like any other contribution: review privacy and accuracy before merging. Setup lives in the repo at .cursor/automations/daily-discord-wiki.md (not loaded by Cursor automatically — activate once at cursor.com/automations/new).

Two ways to contribute

1. Directly on GitHub (no command line)

  1. Open the page you want to edit on GitHub and click the edit (pencil) button, or use the “Suggest an edit” link on the mdBook page. You’ll be prompted to fork the repo — do so.
  2. Make your changes, keeping the markdown clean (see style below).
  3. Commit with a descriptive message.
  4. Open a pull request targeting the master branch.

2. Locally with mdBook (lets you preview)

git clone https://github.com/blivioniag/gfx1030-wiki.git
cd gfx1030-wiki

# Install mdBook (prebuilt binary, no Rust required)
mkdir -p "$HOME/.local/bin"
MDBOOK_VERSION=v0.5.4
curl -sL "https://github.com/rust-lang/mdBook/releases/download/${MDBOOK_VERSION}/mdbook-${MDBOOK_VERSION}-x86_64-unknown-linux-gnu.tar.gz" \
  | tar -xz -C "$HOME/.local/bin"
export PATH="$HOME/.local/bin:$PATH"

mdbook serve   # live preview at http://localhost:3000

Adding a new page

  1. Read Wiki structure — pick the right folder and naming convention.
  2. Create a markdown file under src/<section>/, e.g. src/vllm/my-topic.md.
  3. Add a link in src/SUMMARY.md and the section overview.md if one exists.
  4. Run mdbook build to confirm it renders and there are no broken links.
  5. Update Verification status for benchmarks or community claims.

Markdown style

  • Use ATX headings (#, ##, …) and start each page with a single # title.
  • Wrap shell commands in fenced code blocks with a language hint (```sh).
  • Prefer relative links between wiki pages (e.g. ../vllm/overview.md) so they work locally and deployed.
  • Keep lines readable; hard-wrapping around ~100 columns is fine but not required.

Accuracy

RDNA2 tooling moves quickly. When you add a command, note the ROCm / library version you tested it with if it might matter, and prefer linking to official docs over pasting version-specific numbers that will age.

For pages sourced from Discord or fork release notes, add or update entries in Verification status so readers know what is solid, fork-source, community-reported, or needs verify. Mark unconfirmed throughput claims as community-reported rather than presenting them as wiki-tested facts.

Privacy (Discord → wiki)

This wiki is public on GitHub Pages. When turning Discord messages into docs:

DoDon’t
Summarize operational facts (env vars, topology, bench numbers)Paste message URLs, message IDs, or @mentions
Attribute with channel names (#vllm-rdna) or “community report”Copy Discord usernames, display names, real names, or avatars
Use /path/to/model, ./build/bin/…, -hf org/modelCopy someone’s home directory, hostnames, Linux usernames, or internal IPs
Link public repos (GitHub, Hugging Face, lunnova.dev)Commit bot tokens, guild IDs, or channel IDs
Generalize hardware (“Ice Lake 4× V620 host”, “server board X”)Quote forum thread titles that include a member’s name
Cite motherboard / CPU product models when they help topology advicePaste PCI BDFs (0000:1a:00.0), serials, MAC addresses, or geotags

Public GitHub / Hugging Face org or repo names are fine to link (they are already public attribution). Prefer “community report” / “fork README” over naming Discord people even when the same handle appears on GitHub.

Never commit: DISCORD_BOT_TOKEN, DISCORD_GUILD_ID, .env files, or screenshots of private channels. Cloud Agents should set token + guild ID in Secrets only (see repo README.md).

Automated Discord digest PRs

When the daily digest automation is active, expect occasional draft PRs that only touch src/ (and SUMMARY.md if a page was added). Review them like a human patch:

  • Reject transcripts, @mentions, message links, and personal host/path leaks.
  • Keep unconfirmed tok/s and one-off anecdotes in Verification status as Community / Needs verify.
  • Empty calendar days should produce no PR; do not “fill” the wiki from stale chat.

The agent instructions are .cursor/skills/discord-wiki-digest/SKILL.md. You can still ask a Cloud Agent for an on-demand pass; it should follow the same skill.

Deployment

Merges to master are built and published to GitHub Pages automatically by the .github/workflows/mdbook.yml workflow — no manual deploy step is needed.