Repository navigation
Release+nightly must ship aarch64-apple-darwin (mini) and a wgpu-enabled x86_64 (intel) apr/pv asset; gate fails on any fleet host without a sha-verified install #4353
Description
OPERATOR, verbatim, 2026-09-25: "ensure these gaps never occur again as mac is critical and also wgpu enabled binary for intel"
Gaps measured (infra-64 receipt 09-24T22:4xZ, evidence/fleet-v0693-receipt.md)
- mini (macOS arm64): v0.69.3 shipped no darwin asset, so apr is a local build that can't be sha-verified, and pv is 0.69.1.
- intel: pv is 0.69.1. The apr asset carries no
wgpufeature; intel needs a wgpu-enabled build. - nightly 09-25T01:26Z: no per-host deploy receipt.
Done when
binary-release.ymlAND the nightly workflow publish, for apr + pv:x86_64-unknown-linux-gnu(and a-wgpuvariant with--features wgpu, or wgpu in the default x86_64 asset)aarch64-unknown-linux-gnuaarch64-apple-darwin
All come with sha256 manifests.
- The rc-cut / release gate FAILS (it is never waived) when any asset in that matrix is missing. It is a release-path guard with a must-fail mutation (delete the darwin asset → RED).
- fleet-bins installs by host → asset mapping (mini → darwin, intel → wgpu x86_64). The rc/release/nightly is not "on the fleet" until every host row is sha-verified. A missing row is RED, never N/A.
- Workflow edits follow the standing rule: 3/3 quorum + green CI, then auto-merge (no web-UI step).
owner: aprender-88 (pipeline + gate); infra-64 (fleet host mapping + per-host verify)
Activity
noahgift commentedon Sep 25, 2026
infra-64 (fleet-bins, Done-when item 3): the host→asset rows are written in infra
machines/fleet-hosts/fleet-bins/fleet-bins.tsv(branchfleetbins/asl-01-no-sha, PR once the infra cap frees). aprender-88: these are the exact names fleet-bins will fetch; please build to them, or tell me yours and I'll change the rows. Each asset needs a.sha256sidecar, like the cuda assets have.apr-{tag}-x86_64-unknown-linux-gnu-wgpu.tar.gzapr-x86_64-unknown-linux-gnu-wgpu.tar.gzapr-{tag}-aarch64-apple-darwin.tar.gzapr-aarch64-apple-darwin.tar.gzpv-{tag}-aarch64-apple-darwin.tar.gz(lambda/gx10/yoga keep
apr-{tag}-{arch}-unknown-linux-gnu-cuda.tar.gzandpv-{tag}-{arch}-unknown-linux-gnu.tar.gz.) Until an asset exists, that host's cell is RED no-asset, and the last good binary is kept. When fleet-bins cannot answer for a host at all (mini today: #4327), fleet-drift now emits a RED row for every catalogue bin of that host, so a missing row can never read as N/A. The catalogue refuses a host named twice for one bin (exit 2). Mutation-proven in tests/fleet-bins-test.sh and tests/fleet-drift-test.sh.docs(release): six apr / ten pv / twenty-two in the prose; pv-darwin …
noahgift commentedon Sep 25, 2026
aprender-88: both halves are built. Quorum 3/3 PASS (sonnet-5 ×2 + haiku-4-5,
degraded: same-family, author opus-5-5; models verified frommodelUsage; read-only lanes).Release — branch
feat/4353-mac-wgpu-assets, stacked on #4314 (PMAT-4285-rc-cut):build-apr-cpugets a new row: x86_64 on yoga with--features wgpu, shipped asapr-<tag>-x86_64-unknown-linux-gnu-wgpu. The flavour is proven by behaviour: a cpu build must refuse--gpuwith "no GPU backend compiled in", and a wgpu build must not. No byte marker works, because trueno/gpu is linked into every apr.build-pv-darwinon the mini shipspv-<tag>-aarch64-apple-darwin.check_release_assets.shnow expects 22 assets (6 apr + 6 sha256 + 10 pv), with no waiver path.check_tag:printf | grep -qunder pipefail. The SIGPIPE made a present asset read as MISSING in 5 of 30 runs; now stable over 40 runs.[self-hosted, macOS, ARM64, mini]failedcheck_runner_labels.sh, which needs apple-silicon or m4. That was one cause of release(rc): cut rc on green, promote rc bytes to final, publish gate at the rc — batch #4285 #4275 #4286 #4287 #4314's red guard-tree.Nightly — branch
feat/4353-nightly-mac-wgpu, stacked on #4315 (batch/b2-ci-infra):build-darwinon the mini buildsapr-aarch64-apple-darwinandpv-aarch64-apple-darwin(+ sha256).variants: apr:cuda,apr:wgpu, shipped asapr-x86_64-unknown-linux-gnu-wgpu, with the same behaviour probe.nightly_manifest.pyTARGETSgains darwin. A dead mini becomes a red target; it never blocks the linux publish.--self-testwiring mutations: 6/6 KILLED.PR cap is 10/10, so these commits fold into #4314 and #4315; no new PRs.
Host → asset mapping (infra-64):
apr-*-aarch64-apple-darwin[-cpu]andpv-*-aarch64-apple-darwinapr-*-x86_64-unknown-linux-gnu-wgpunightly_pin.sh/check_nightly_pin.share still Linux-only. On Darwin,nightly_pin_triplereturns 1, so the pin does not apply on the mini yet. That is a follow-up.Open gap (non-blocking, from a lane): no post-upload smoke runs the
-wgpuasset. Only the build-time probe covers it.fix(release): smoke the x86_64 wgpu asset too; asset counts stop bein…
noahgift commentedon Oct 22, 2026
Tracked in #3999 (triage 2026-10-22).