action

launchbound safety gate — GitHub Action

Run the convergence safety gate over a cuda-oxide kernel's tuning space on every push, with no GPU anywhere: every launch/specialization candidate comes back admitted, admitted-with-caveats, or refused with a rule ID and source span. This is the part of the tuning pipeline that finds the bugs, and it runs on a plain runner.

Paste this

name: convergence-gate
on: [push, pull_request]

permissions:
  contents: read

jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      # Your repo, containing the kernel crate + kernel.toml.
      - uses: actions/checkout@v4
        with:
          path: myrepo
      # The kernel's cuda-device path dependency must resolve: check out the
      # cuda-oxide commit your crate pins, in the layout your Cargo.toml
      # expects (here: a sibling directory).
      - uses: actions/checkout@v4
        with:
          repository: NVlabs/cuda-oxide
          ref: 50d07314eb8b7d5ec821ba02b0048a753c20dd4e
          path: cuda-oxide

      - name: Safety gate
        id: gate
        uses: vyncint/launchbound/action@v1
        with:
          kernel: myrepo/kernels/my-reduction
          cc: "8.6"          # the part you will actually run on
          fail-on: refused   # keep this space all-clean

      - name: Use the outputs
        run: |
          echo "verdict: ${{ steps.gate.outputs.verdict }}"
          echo "${{ steps.gate.outputs.refused }} refused of \
          $(( ${{ steps.gate.outputs.admitted }} + ${{ steps.gate.outputs.caveats }} + ${{ steps.gate.outputs.refused }} )) candidates"

What it needs from your kernel

The kernel directory must contain a kernel.toml declaring the tuning space (see the corpus for six worked examples) and a device-only crate whose cuda-device path dependency resolves in the checkout layout above. src/params.rs carries the compile-time parameters the gate specializes per candidate.

Inputs

input	default
kernel	—	path to the kernel directory (required)
cc	—	target compute capability, e.g. "8.6" (required; verdicts do not transfer across parts)
fail-on	tool-error	never, refused, or tool-error
version	latest	launchbound-cli release to install
reconverge-version	0.1.11	reconverge release from crates.io — moves in lockstep with toolchain
toolchain	nightly-2026-04-03	the nightly that built that reconverge
summary	"true"	write the verdict table to the job summary

Outputs

verdict (clean / caveats / refused / tool-error), the counts (admitted, caveats, refused, tool-errors), and json — the full gate record with rule IDs and source spans, for anything the counts miss.

The toolchain contract

The gate compiles and analyzes your kernel under the action's toolchain input, not your repository's rust-toolchain.toml — that is inherent to reconverge being a rustc driver that must pair with one exact nightly. If your kernel needs a different nightly, pass a matching toolchain/reconverge-version pair; changing one without the other fails at install time.

Honesty notes

A clean gate is not a proof of correctness — the gate inherits reconverge's documented limits, and the launch-shape classifier recognizes the measured warp_id() family (see docs/LIMITATIONS.md). The gate covers CUDA kernels only; there is no MSL analyzer, so Metal code is not checked. The first run builds reconverge from its crates.io release (a few minutes); nothing is cached, so the gate can never be stale, poisoned, or the wrong version.