[drakestapleton]

Purpose and scope

Refs #189. ADR 0043 leaves the imported C runtime frozen and keeps the versioned public API in this repo. The host gate for that surface is make test-runtime-api.

That command was already a CI job before this PR. .github/workflows/runtime-api.yml (added in 342db51) runs it on every pull request and on pushes to main, on GitHub-hosted ubuntu-24.04-arm, which has no GB10. The last green run on main before this branch is Actions run 41073733502 at 522ee0d (RESULT: PASS).

This PR only changes that workflow file:

• The job still runs make test-runtime-api, and it fails unless the log ends with RESULT: PASS (...).
• A following step prints GB10 sm_121 chip execution: NOT_RUN. The job does not open /dev/nvidia0, /dev/nvidia-uvm or /dev/aien-gpu, and a host pass is not a chip pass.
• make test-runtime-api-semver still compares the manifest with the pull request base (or the previous commit on a main push).

No frozen C under ADR 0043, no src/omega_cubin.*, no om/cubin/, and no file from PR #170.

How this sits with #90 and #196

This stays a separate workflow. It does not edit ci.yml or host.yml.

• [PUBLIC-READY P0] Reproduce host gates from fresh checkout and enforce them on PRs #90's host-gates work owns make test in the existing arm64 host-tests job. That suite already lists test-runtime-api as one row. It does not run test-runtime-api-semver against the pull request base.
• CI: add an x86-64 host-tests job (portable suite + freestanding link) #196's x86-64 job owns the portable host suite and the freestanding link on ubuntu-latest. When that job runs make test, it will run test-runtime-api again on x86-64. It will not run the semver-against-base step.
• This job is the short manifest, negative-control and semver gate. Local x86-64 confirmation of the same make test-runtime-api is below. A duplicate x86 job is left to CI: add an x86-64 host-tests job (portable suite + freestanding link) #196.

Evidence on the exact commit

git rev-parse HEAD: b3609380179b96c0d2eb6ec33026edceb21e2bff (parent 522ee0d).

PR job: runtime-api on ubuntu-24.04-arm, pass in 56s. Log lines:

• RESULT: PASS (runtime API 0.6.0: 295 functions, 45 structs, 420 fields, 211 constants match the headers; surface 698 names, 192 INTERNAL)
• RESULT: PASS (every negative control was caught)
• GB10 sm_121 chip execution: NOT_RUN (host manifest check only; /dev/nvidia0, /dev/nvidia-uvm and /dev/aien-gpu stay closed)
• RESULT: PASS (version 0.6.0 -> 0.6.0; the entry changes need none)

Same commands on this VM before push: Linux x86_64, gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0, /dev/nvidia0 and /dev/aien-gpu absent. make test-runtime-api exit 0 with the same two RESULT: PASS lines. make test-runtime-api-semver RUNTIME_API_BASE=HEAD exit 0 with RESULT: PASS (version 0.6.0 -> 0.6.0; the entry changes need none). No new receipt folder: the check prints its own result line.

• make test-runtime-api (host-only) — PASS, locally and in the PR job above.
• make test-runtime-api-semver against the base — PASS in the PR job (0.6.0 -> 0.6.0, no entry change).
• make test (host-only) — NOT RUN. Unchanged by this diff; still the host-tests job in ci.yml.
• make test-native-artifact (host-only) — NOT RUN. Out of scope.
• make zero-cuda-gate (ELF audit and negative control) — NOT RUN. Out of scope.
• build/hello_gpu --sim (simulator only) — NOT RUN. Out of scope.
• make doctor (diagnostic; NOT READY on unsupported machines is expected) — NOT RUN. Out of scope.
• GB10 hardware tests, only if an authorized GPU window is available (otherwise NOT RUN) — NOT RUN. The workflow step reports that. This VM has no GB10 and no /dev/aien-gpu.

Correctness, compatibility and provenance

• Negative controls and affected contracts are documented. The negative controls are the existing tools/runtime_api_negative.sh run inside make test-runtime-api. The new check is that a zero exit still ends on RESULT: PASS.
• API/ABI and Omega consumer effects are identified; source and runtime status claims are not overstated. No header, manifest or ABI change. The chip path is NOT_RUN, not PASS.
• Third-party material, licensing, and public-history safety were checked for additions. Workflow comment and shell only.
• No secrets, personal machine paths, private addresses, other processes' logs, vendor binaries, or unauthorized generated artifacts were included.
• Existing receipts remain immutable; new results have their own evidence folder. No receipt was edited. None was added.

Known limitations

The job passes the host manifest check. It does not qualify the runtime on a GB10. No source failure turned up, so there is no candidate follow-up issue.

— Grokbot

  

[drakestapleton]

Review at b360938: ready to merge once CI is green. At review time only library had passed; runtime-api (run 41074758124) was in progress and everything else was queued.

Scope. There's one file, .github/workflows/runtime-api.yml (+18/−3), and it stays inside the existing runtime-api job.

• L2-6: header comment.
• L25-35: the existing step is wrapped so the exit code is kept and the last line has to match RESULT: PASS (…).
• L36-38: a new print-only step.

Triggers, runner, permissions, checkout pin and the semver step are unchanged. ci.yml and host.yml are untouched, so this is no rewrite.

NOT_RUN and fake-green check

• L25-35 is stricter than before. A non-zero rc is propagated (L30), and an exit 0 without a final PASS line fails (L33-34). There's no continue-on-error, no || true and no if: skip. || rc=$? at L28 is captured and then re-raised.
• L36-38 only prints NOT_RUN. It's honest, but it shows as a green step named "GB10 sm_121 chip execution" in the Checks UI. Optional: rename it to something like "GB10 sm_121 chip execution (NOT_RUN)", or write the line to $GITHUB_STEP_SUMMARY, so a green tick isn't read as a chip pass.
• L46-47 is pre-existing and outside this diff: if the base has no manifest, the step exits 0 and prints NOT_RUN. That's a host skip, not a chip claim, so it's fine as is.
• The body makes no GB10 behavior claim. The evidence is an x86_64 host run only, with GB10 marked NOT_RUN. That matches #189's done-when.

Fit with #90 and #196. There's no shared file or job, and the check name runtime-api is unique. I'm assuming #90 edits ci.yml host-tests and #196 adds an x86 job, which both need to be named apart from host-tests and runtime-api. The only coupling is the L5 comment pointing at the #196 job. Suggested merge order: this PR, then #196, then #90. This PR is independent and smallest; #196 and #90 are both likely to touch ci.yml, so whichever merges second rebases.

Follow-ups. The body lists none: "No source failure turned up, so there is no candidate follow-up issue."

— Grokbot

[drakestapleton]

Re-review at d71ac1c: ready to merge.

• Fast-forward. Comparing b360938 with d71ac1c shows ahead_by 1, behind_by 0, so the branch wasn't force-pushed. The new commit only touches .github/workflows/runtime-api.yml, +2/-2:

  • L36: the step is renamed to GB10 sm_121 chip execution (NOT_RUN).
  • L38: the same printf NOT_RUN line now goes through | tee -a "$GITHUB_STEP_SUMMARY".

  Nothing else changed. The PR's recorded base is still 522ee0d. Main has since moved with #199, but GitHub reports the PR as mergeable clean, with no conflicts.

• CI. All 12 checks at d71ac1c are green: runtime-api run 41075608687 (job 124130836358; its head_sha is d71ac1c), ci run 41075608852, hello-sim 41075608714, library 41075608711, drift 41075608712, and CodeQL 41075599653.

• Log. In job 124130836358, step 5 GB10 sm_121 chip execution (NOT_RUN) concluded success. At 14:57:32Z (9:57 AM CT) the log printed GB10 sm_121 chip execution: NOT_RUN (host manifest check only; /dev/nvidia0, /dev/nvidia-uvm and /dev/aien-gpu stay closed) under shell: /usr/bin/bash -e {0}, and there was no error. tee is the last command in the pipe, so if the summary write failed, the step would fail.

• Fake-green check. It's clean: no continue-on-error, || true or if: skip, and RESULT: PASS is still enforced by the case at L33-35. GB10 stays NOT_RUN, and the PR makes no chip claim.

— Grokbot