b70-optimization-lab

Community Contributions

This tree holds work contributed from outside the reference lab. It exists so that contributed evidence can be preserved and credited without being mixed into the promoted ledger before it has been checked here.

Nothing in this directory is a lab result. A recipe here may be correct, partially correct, hardware-specific, or wrong. Read the STATUS.md of an entry before running anything from it.

Boundary Rule

Directory Contains Who may add
results/ Promoted, closed-out result packets Maintainer, B70-tested or better
repro/ Copy-ready recipes for promoted results Maintainer, B70-tested or better
community/ Contributed work at any evidence level Anyone

A contribution enters community/ first. It moves into repro/ or results/ only after it reaches B70-tested or higher in this lab, and the move is a separate maintainer commit that records what was actually run. A community-reported entry never enters those directories, however useful it is.

This boundary is about evidence provenance, not contributor trust. A careful contributor with no B70 access produces community-reported work by definition, because the reference hardware is here.

Entry Layout

community/<contributor>-<model>-<topic>/
  STATUS.md      required; evidence label, provenance, what was tested here
  README.md      the contribution as submitted
  validation/    logs, JSON, and commands from any local validation attempt

STATUS.md is mandatory and is the first file a reader should open. Copy STATUS-TEMPLATE.md to start one. An entry without a current STATUS.md is incomplete regardless of how good its README is.

The contribution README.md is preserved as submitted. Corrections belong in STATUS.md under “Known Issues”, not silently edited into the contributor’s text. If the recipe is later fixed, record the fix and its author explicitly.

One exception: a maintainer note may be added at the top of a contributed README, clearly labelled as such and separated from the contributor’s text by a horizontal rule. Its purpose is to stop a reader who lands directly on the file from mistaking a pending submission for a lab result, and to point at STATUS.md. It states status and directs the reader; it does not argue with the contribution.

Evidence Labels

Labels are defined in docs/contribution-verification.md and are used unchanged here: community-reported, B70-tested, B70-verified, matching-hardware verified, invalid, superseded.

Two fields are tracked separately in every STATUS.md and must not be collapsed into one:

A contribution can be safe to read and merge while its performance claim remains entirely unverified. That is the normal state of a new entry.

Validation Procedure

Local validation follows the eight-step procedure in docs/contribution-verification.md. Two points matter most in practice and are restated here:

  1. Protect current work. CURRENT.md is authoritative for active lanes. Contributed work is validated only when the reference GPUs are free, in isolated trees, ports, and result directories, and never by disturbing an active runtime tree.
  2. Record negative results. A contribution that fails to run here is evidence, not a non-event. Keep the logs in validation/ and say plainly in STATUS.md what failed, on what host identity, and what remains unknown.

When a validation attempt is blocked by local infrastructure rather than by the contribution itself, say so explicitly. “Not tested here because the box lacked a container runtime” and “tested here and failed” are different findings and must not be recorded the same way.

Index

Entry Contributor Source Evidence level Tested here
Qwen3.6 27B FP8 TP2 Docker dominick253 PR #9 B70-tested Yes; recipe runs at 30.171 tok/s median decode, inside the reported range