Issues / #909

#909 Proposal (docs only): a short "contributing a change or report" section for AGENTS.md, and a pinned list of test requests. Yes/no?

open · @paulhothersall · 1 comentários · No GitHub

BenchmarksServer & APIMulti-GPUAMD / HIPNVIDIA / CUDADocumentation

Descrição

**TL;DR.** The repo has 223 open PRs and 104 open issues, and AGENTS.md has no section on contributing. I can send a docs-only PR (no engine code) with a short contributor section for `AGENTS.md`, written from your own review comments below, and keep a small "test requests / hardware wanted" list. Three independent yes/no questions at the end; "no" or silence costs you nothing.

(Disclosure: I drafted this with an AI assistant, Claude Code, and checked every reference against the repo myself. I am a contributor with a Tesla P40 + RTX 3070 box, not a maintainer. #875 and #880 are examples of the reports I send.)

## What your reviews already say (please correct me)
- **One change per PR** (#390 asked for the carve, the helper tier and the Monitor changes as separate PRs).
- **Opt-in, and the default path stays byte-identical** (#492 and #490 review; #598 / #663: "output byte-identical", "we keep the default byte-identical").
- **Build on every backend a file touches** (#492 review: the HIP compat header lacked `cudaPointerGetAttributes`; #531: the HIP shim gained the peer-access calls).
- **Defaults change only where it was measured faster** (#583: the byte-budget ring is used only where it won, with the 4K numbers).
- **Reports from hardware you do not have are welcome and used:** #639 ("neither of us has NVIDIA multi-GPU"; asked for a re-run on the MI50s), #663 ("We have one GPU, so the 2x 3090 numbers are yours").
- **Community benchmark reports** follow `docs/COMMUNITY_BENCHMARKS.md`, and AGENTS.md already says: "measured numbers with what they were measured on, no claims without a measurement."

## Proposal
1. A short "Contributing a change or report" section in `AGENTS.md` (about ten lines, linking `docs/COMMUNITY_BENCHMARKS.md`): the five points above, in your words.
2. A short pinned list of test requests, for example "needs a run on sm_61", "needs a 32 GB host", "needs a non-AVX2 CPU". I can take items that match my hardware and report in the benchmark format.
3. Templates: PR #401 (Joe-Huber, open since 10-01) already proposes issue and PR templates. I would not duplicate it. If useful, I can suggest to that PR a few extra fields for hardware reports (GPU + PCIe width, CPU + RAM speed/channels, build commit, "what I did not test") and for PRs ("default path unchanged? how checked", "builds on: CUDA / HIP / SYCL", "measured on").

## Questions
1. Add the contributor section to `AGENTS.md` (I send exactly the text, you edit)?
2. Keep a pinned "test requests / hardware wanted" list?
3. Should I suggest the extra template fields on #401, or leave templates to you?

No site

Links install, modelos, releases.