Skip to content

Set the approved Phase 8.4 release tuple versions - #104

Open
perNyfelt wants to merge 2 commits into
mainfrom
phase8-4-wp3-release-versions
Open

perNyfelt wants to merge 2 commits into
mainfrom
phase8-4-wp3-release-versions

Conversation

@perNyfelt

@perNyfelt perNyfelt commented Nov 17, 2026 •

Copy link
Copy Markdown
Member

The owner approved the audited 8.4 tuple on 2026-11-17 (recorded in
req/phase8-4-release-checklist.md). Per the Phase 8.4 plan's WP3, the release
versions ride a separate reviewed PR: the candidate tag jmlx-phase8.4-rc.N
anchors at this change's main merge SHA, and the release runs from a clean
detached checkout of that tag.

Changes

  • Nine module versions — jinja 0.6.2, tokenizer 0.1.2, vision 0.1.2,
    native-macos-arm64 0.1.0, ffi 0.5.0, core 0.5.0, models 0.1.0,
    spring-ai 0.1.0, langchain4j 0.1.0 — exactly the tuple in
    tools/release-smoke/released-8.4.json (no reused coordinates; every entry
    is local, so there is nothing to restore).
  • Smoke CI defaults move to the same versions, the documented same-PR rule
    (tools/release-smoke/gradle.properties: CI mode resolves the defaults
    unstripped, so publishing producers and the smoke consumer agree).
  • Changelogs finalized (WP3): jinja and tokenizer gain dated 0.6.2 /
    0.1.2 sections recording the non-functional differences only (.module
    Gradle 9.7.0 to 9.9.0; tokenizer additionally Jackson 3.1.2 to 3.3.0 and its
    jinja dependency 0.6.1 to 0.6.2); core 0.5.0 and models 0.1.0 roll their
    Unreleased entries in; the two development-cycle status notes are dropped.
    The section dates are this PR's date, not the (future) publication day;
    adjust them before tagging if the merge or publication lands later.
  • selftest e2e preflight case fixed: the 8.4 preflight reads the actual
    build.gradle files, so on a release-ready tree the version validation
    passes and the case no longer triggers its intended failure — it falls into
    the publish path (side effects) and fails downstream. The case now forces
    its failure with a JMLX_JINJA_VERSION=0.7.0 override, deterministic on any
    tree state, and greps the preflight's own override message.
  • Doc coordinates: the adapter READMEs no longer name -SNAPSHOT
    development coordinates, and the hf-reference README's local jar names use
    the new core/ffi versions.

Notes

  • AGENTS.md's version paragraph is deliberately untouched: the 8.1 precedent
    moves it to the next development line in the post-publication dev-bump PR.
  • The checklist's "Approve the audited tuple" gate carries the dated approval
    note; the box stays open until the audit rerun at the final candidate SHA.
  • The root project version (unpublished examples/benchmarks) remains
    0.5.0-SNAPSHOT.

Validation

  • ./tools/release-smoke/selftest.sh — passes (50 cases, including the fixed
    e2e preflight case).
  • ./tools/release-smoke/run.sh ci --profile pure-java — publishes the three
    pure-Java modules at 0.6.2/0.1.2/0.1.2; the Java 21 consumer resolves
    exactly those coordinates; POM/module agreement, license, JPMS, NOTICE,
    javadoc and sources verified; pure smoke passed.
  • ./tools/release-smoke/run.sh ci — publishes all nine modules; the Java 25
    consumer resolves all nine new coordinates; POM-only and module-metadata
    runtime classpaths agree; the adapter smoke (chat, streaming, embeddings,
    cancellation, virtual threads, close) and the seeded two-request batch both
    pass.
  • ./releaseAll.sh --dry-run (Temurin 25.0.4.1 + 21.0.13 homes) — the plan's
    selected set matches the released-8.4.json tuple exactly (nine modules, no
    SNAPSHOTs); signing reported warning-only (no key in this environment).

Next steps after merge

  1. Anchor jmlx-phase8.4-rc.1 at the merge SHA; work from a clean detached
    checkout of the tag.
  2. ./releaseAll.sh --bundle-only with JMLX_RELEASE_CANDIDATE_TAG.
  3. Consumer bundle smokes (macos-full, macos-pure, local linux-pure) and
    the capacity measurement against the preliminary record.
  4. release-candidate.yml dispatch in the reviewed minimum-host window, final
    receipt, then explicit publication authorization.

CI fix (run 44181118435)

The first CI run had two failures; they are unrelated to each other.

1. java job: "Verify release isolation and task evidence" (real, fixed in commit 2dd1223).
Moving jmlx-native-macos-arm64 off -SNAPSHOT activated the version-conditional publish
dependency on verifyNativeReleasePayload (an Exec task running
python3 release_inputs.py native ...). On an unstaged checkout the Exec task fails before
the publish task's unconditional doFirst refusal runs, so the release selftest's
combined-staging fixture — which copies the real build.gradle into an unstaged temporary
root — saw a python "can't open file" failure instead of the canonical
"Refusing to publish jmlx-native-macos-arm64 without staged native binaries" message. This
also affected the real-world message on an unstaged release-version checkout. Fix: the payload
check now skips via onlyIf { hasValidatedStagedNativeFiles() }, mirroring
verifyPackagedNativeResources's existing lenient pattern; with staging missing, the publish
task's own hard, unconditional doFirst gate fires the canonical message; with staging valid
but the payload wrong, the same task still fails the release.
Validation: the fixture failure reproduced locally against the committed build.gradle; with the
fix the publish task fails with the canonical message and stages no jar;
bash tools/release-all/selftest.sh passes 84/84; verifyNativeReleasePayload still runs and
passes on this staged machine.

2. native job: Phase71ContractsTest > repeatedForwardDoesNotRetainActivationBuffers (flake, not this PR).
Line 249 is an MLXMemory.activeBytes() <= baseline + 65536 Metal-allocator tolerance
assertion — the same flake class as the known hosted GatedDeltaMemoryTest pooled-buffer
over-charge on macos-26 runners. This PR changes no test and no native code, and the same
:jmlx-core:test suite passed on this hosted infrastructure in PR #102's green merged-head
run (44167458671). Expected to pass on the new run.

The owner approved the audited tuple 2026-11-17 (recorded in
req/phase8-4-release-checklist.md). Per the Phase 8.4 plan's WP3, the
release versions ride a separate reviewed PR: the candidate tag anchors
at this change's main merge SHA and the release runs from a clean
detached checkout of that tag.

- Set the nine module versions -- jinja 0.6.2, tokenizer 0.1.2,
  vision 0.1.2, native-macos-arm64 0.1.0, ffi 0.5.0, core 0.5.0,
  models 0.1.0, spring-ai 0.1.0 and langchain4j 0.1.0 -- exactly the
  tuple in tools/release-smoke/released-8.4.json.
- Move the smoke consumer's CI defaults to the same versions, the
  documented same-PR rule: CI mode resolves the defaults unstripped, so
  the consumer and the publishing producers agree.
- Finalize the changelogs: dated 0.6.2 and 0.1.2 sections for jinja and
  tokenizer (no functional changes; .module and dependency differences
  only), core 0.5.0 and models 0.1.0 roll their Unreleased entries, and
  the two development-cycle status notes are dropped.
- The selftest e2e preflight case now forces its failure with a
  JMLX_JINJA_VERSION override: the 8.4 preflight reads the actual
  build.gradle files, so a release-ready tree passes version validation
  on its own and would run the case past the preflight.
- Refresh the coordinate references in the adapter READMEs and the
  hf-reference README's local jar names.
The version change in this PR moves jmlx-native-macos-arm64 off
-SNAPSHOT, which activates the release-version publish dependency on
verifyNativeReleasePayload. On an unstaged checkout that Exec task
now fails first (a python error, or a missing bootstrap script in a
partial checkout) before the publish task's unconditional doFirst
refusal runs, so the canonical message telling the human to run
bootstrap-native.sh never prints. The release selftest's
combined-staging fixture caught this in CI (run 44181118435).

The payload check now skips via onlyIf while native/install/lib is
unstaged, mirroring verifyPackagedNativeResources' existing lenient
pattern: with no staged files there is no real payload to check, and
the publish task's own hard, unconditional doFirst gate fires with
the canonical message instead. Staging valid but the payload wrong
still fails the release in the same task.
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.

1 participant