chore(local-inference): bump fork gitlink to 2bdcef890 — portable non-Apple kokoro fast path (thread pool + NEON)

[lalalune]

Follow-up to #11517 (squash-merged before this landed): bumps plugins/plugin-local-inference/native/llama.cpp from 114eee08e (Accelerate-only) to 2bdcef890 — a linear superset (2bdcef890 ⊃ 114eee08e Accelerate ⊃ ae0a1eed2 converter) adding the portable non-Apple (Android/Windows/Linux) kokoro fast path: internal thread pool over output channels/gate rows + aarch64 NEON AXPY/dot MACs for conv1d/convtranspose1d/linear/LSTM gates, with the scalar loops retained as reference (KOKORO_FORCE_SCALAR).

Evidence (.github/issue-evidence/kokoro-metal-perf/): portable-fast-parity-microbench.log + README section — 23/23 parity checks vs scalar (max |Δ| 1.9e-5 < 1e-4) on real generator shapes; conv up to 107×, convtranspose 9–19×, LSTM 4.6× vs scalar; NDK aarch64/x86_64 Android cross-compile clean; Apple Accelerate path rebuilt green.

Refs #9033.

🤖 Generated with Claude Code

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 2ba14516-eb9c-49c5-b9ed-545ddbdbd16c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch chore/local-inference-gitlink-2bdcef890-portable-kokoro

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[lalalune]

Closing as superseded by #11584 — the exact bump is already on develop; verification of the pointer itself passed

Develop's gitlink was moved to the identical target by 9426fbf75f9 ("chore(local-inference): bump llama.cpp gitlink to 2bdcef890 — portable non-Apple kokoro fast path (#11584)", merged ~17:00 EDT today), and both evidence files (.github/issue-evidence/kokoro-metal-perf/README.md + portable-fast-parity-microbench.log) are on develop too. Nothing in this PR remains to land.

The pointer verification (done here, applies equally to the landed #11584):

• 2bdcef890ef7c153b9c021076814162ca2caa340 exists on the public fork: refs/heads/kokoro-portable-fast, refs/heads/perf/kokoro-accelerate-blas, refs/pull/38/head (verified via git ls-remote https://github.com/elizaOS/llama.cpp.git).
• git merge-base --is-ancestor 114eee08e… 2bdcef890… → descends from the feat(local-inference): Gemma-4 cutover finish — vision + kokoro 417× + gemma4-assistant MTP + stale-model safety #11517 pin; no pointer regression.
• Delta 114eee08e..2bdcef890 is a single commit touching only tools/kokoro/** (+650/−7). The 7 removed lines are a comment block plus #if defined(__APPLE__) → #if defined(__APPLE__) && !defined(KOKORO_NO_ACCELERATE) && !defined(KOKORO_FORCE_SCALAR) — Apple Accelerate remains the default; the new macros are opt-out knobs used by the parity tests. Everything else is additive (portable thread-pool + NEON path, three new test targets).
• Pushed a protective pin tag on the fork so the SHA survives branch rewinds: eliza-pin-11545-kokoro-portable → 2bdcef890… (joins eliza-pin-11517-gemma4-mtp-kokoro-blas → 114eee08e…).

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.