House rules
- Prefer llama.cpp over vLLM for consumers. Prefer Q5 over Q4. Never below Q4 — including draft and MTP.
- Models from NVMe, not HDD. One LLM per GPU unless you deliberately combine.
- Do not reintroduce Qwen2.5 or Qwen3-8B. Abliterated stays educational.
- Do not invent flags, VRAM, RAM, or issue numbers. Mark gaps
TODO. - Official vs works: official means the SKU is on AMD’s current matrix or matching notes.
HSA_OVERRIDE_GFX_VERSIONis unofficial. Never present it as a plan.
Edit data, not mythology
Most living facts live in YAML under src/data/:
rocm-problems.yaml— this tracker. Every row needs a last-checked date and an https source.errors.yaml— troubleshooter patterns.stack-updates.yaml— alerts matched to saved rigs.gpus.yaml,stacks.yaml,what-works.yaml.
Seed only what you verified. If you cannot open the GitHub issue, leave the row out or mark todo: true.
Keep main green
ACCOUNTS_ENABLED=0 FORUM_ENABLED=0 npm run buildmust stay a static site.- Do not import
src/server/*from a page that prerenders when flags are off, except behindif (accountsEnabled())/forumEnabled()dynamic import. - Add tests next to the lib you changed.
npm testshould stay green.
Voice
Plain English. No lorem. No “welcome to your app.” Cite with a real URL. Affiliate links are Amazon Associates, Rakuten, and B&H only — seedisclosure.
Translations
A translation is welcome if it does not drift from the English sources and keeps every TODO. Do not translate an invented number into confidence. ROCm’s own documentation translations belong inROCm/ROCm, not here.
Conduct
The forum code of conduct applies to comments on this site too. Argue the log. Abliterated is educational, not the default stack.