Skip to content

chore: add issue forms, a PR template and CODEOWNERS - #93

Merged
vyncint merged 1 commit into
mainfrom
chore/issue-and-pr-templates
Sep 22, 2026
Merged

vyncint merged 1 commit into
mainfrom
chore/issue-and-pr-templates

Conversation

@vyncint

@vyncint vyncint commented Sep 22, 2026

Copy link
Copy Markdown
Owner

.github/ held scripts/ and workflows/ and nothing else. No ISSUE_TEMPLATE/, no pull_request_template.md, no CODEOWNERS. The ecosystem drift log has carried "launchbound issue templates … no template directory observed" as a deferred item since 2026-09-07.

Why forms rather than free text

More than for a library, because of what a report here has to contain before anyone can act on it. A result depends on the GPU, the driver version, the CUDA version, the compute capability, the cuda-oxide commit, the nightly, and which of the three backends produced the number — docs/LIMITATIONS.md has a whole section ("Results do not port") on precisely that. "The tuner picked a slow configuration" is the start of a conversation; the same report with its provenance block is a bug.

Three forms:

  • The gate admitted or refused the wrong configuration. The one that matters most — refusing a fast configuration is the product, so getting it wrong in either direction is the product failing. It asks which direction, and for evidence: a hang with a driver version, a compute-sanitizer synccheck report, a lane argument.
  • A measurement looks wrong. Asks for the run directory, which already carries the device, driver, warmup and repeat counts and every sample. Its preamble states the two things people file against without knowing: results do not port between capabilities, and overlapping intervals are deliberately not ranked.
  • An admitted candidate failed to build. The hole docs/LIMITATIONS.md already documents under "cuda-oxide is alpha" — cargo check under the reconverge driver does not evaluate all codegen-time consts. Each report names another one.

Blank issues are off, and the config links a wrong kernel verdict to reconverge, which this project ships as a component and does not reimplement.

The PR checklist carries the two things a contributor here can only learn by being told: a change that moves a measured number needs a bench: changelog entry naming the hardware, and the lockstep pin set moves in a commit of its own.

Closes #80

`.github/` held scripts/ and workflows/ and nothing else, and the
ecosystem drift log has carried this as a deferred item since 2026-09-07.

The forms matter more here than they would for a library. A result
depends on the GPU, the driver, the CUDA version, the capability, the
cuda-oxide commit, the nightly, and which of the three backends produced
the number -- docs/LIMITATIONS.md has a whole section on that -- so a
free-text "the tuner picked a slow configuration" is unactionable, and
the same report with its provenance block is a bug.

Three forms: a wrong convergence verdict in either direction, which is
the report that matters most since refusing a fast configuration IS the
product; a measurement that disagrees with what you see; and the known
hole where a gate-clean candidate still fails the real compile. Blank
issues off. A finding wrong about the kernel rather than the launch shape
is pointed at reconverge, which this project ships and does not
reimplement.

Signed-off-by: Vyncint Ng <chivy.nguyen@manabie.com>
@vyncint
vyncint merged commit 56481fc into main Sep 22, 2026
11 checks passed
@vyncint
vyncint deleted the chore/issue-and-pr-templates branch September 22, 2026 06:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

No issue templates, PR template or CODEOWNERS

2 participants