[leehack]

Qualifies upstream stable v0.8.0 (8642541b, b11912 + 48 commits) as the next native runtime, for llamadart 0.17.1 (#112, plan llamadart#922). Supersedes #125 (b11912): its four commits are carried unchanged onto current main, plus one retarget commit.

Draft, not for merge. No tag, release, workflow dispatch or downstream pin change. The production submodule stays at v0.7.0.

Merging this publishes v0.8.0. The scheduled stable dispatcher already selects v0.8.0. Its 2026-10-29 dispatch from main failed in every build job because main's wrapper does not compile against the v0.8.0 mtmd API, so nothing was published. Once this is on main, the next scheduled run builds and publishes v0.8.0 as latest with no further code change. Merge only when the maintainer approves publication.

Changes

• validate_wrapper.yml: a v0.8.0 row in every native lane, pinned to the exact SHA (replaces [0.17] Qualify llama.cpp b11912 as the next runtime #125's b11912 rows). test_ci_scope.py covers it.
• Wrapper (from [0.17] Qualify llama.cpp b11912 as the next runtime #125): upstream #30902 gives each TTS generation its own mtmd_gen_state; the Seed every Qwen3-TTS codebook from the request on each synthesis #122 priming call compiles only without that API. Exports unchanged (89). TTS still accepts only Qwen3-TTS; Soprano and LFM3-Audio report unsupported.
• Android ISA: v0.8.0 ggml/src fingerprint bound to the unchanged KleidiAI v1.32.0 kai digest and the b11912 function set. The only ggml-cpu change since b11912 is the non-contiguous scale op (#31146).
• Docs: docs/v080_qualification.md (patch status, consumer ABI delta from v0.7.0), docs/v080_android_isa_qualification.md.

Carried patches

All three apply strictly (llama_cpp_patches.py apply --strict).

• 0001 (Adreno 750, Investigate Galaxy S24 Vulkan crash during first GGUF generation #79): still needed. Upstream #29165 is an open draft; the matvec reduction-mode selection is identical from b11256 to v0.8.0.
• 0002 (buffer device address, Vulkan may crash on Mali-G68: null vkGetBufferDeviceAddress in ggml_vk_create_buffer llamadart#782): still needed; v0.8.0 still trusts bufferDeviceAddress without the entry-point check.
• 0003 (Qualify and repair the v0.7.0 CPU crash after enabling context embeddings #114): redundant; #31086 re-reserves on every toggle. A v0.8.0 build without 0003 passes 57/57 native tests, including model-session and model-embedding-toggle. It cannot leave series.json here: the pinned v0.7.0 and its rows build from the same series and need it. Remove it with the pin move to v0.8.0.

Local evidence (macOS arm64, M4 Max)

Check	Result
Wrapper contract build (Debug, CI flags) + ctest, v0.8.0 with patches	57/57
#114 patchless build (Release, CPU, Metal/KleidiAI off, 0003 removed) + ctest	57/57
validate_exports.py / grammar boundary	89 exports / 2000 accepted, 2001 rejected
Android armv8.2 ISA (NDK 28.2)	PASS: 261,765 instructions; 3,130 scalable, all in 22 exact functions
KleidiAI dispatch test, SME on and off	128 masks + 12 matmuls pass
Python suite	279 OK, 1 skip; *_test.py 22 OK
actionlint	clean except the repo's known windows-11-vs2026-arm label
Qwen3-TTS 1.7B Q4_K_M + Q8_0 projector tts_smoke (CPU, model hashes as in #125)	PASS: same seed repeats PCM, distinct seeds differ, cancel checks pass

Not covered

• No lane executes the new SVE Q8 kernels (needs SVE256 + SVE I8MM).
• Model, family and device qualification (Investigate Galaxy S24 Vulkan crash during first GGUF generation #79 on S24, Qualify Intel Mac AMD Metal output correctness and fail-safe selection #116 on Intel/AMD) stays in Qualify post-v0.6.0 upstream correctness fixes and consumer compatibility #112.
• High-risk independent audit not yet run.
• Linux x64 CUDA release builds fail on ubuntu-latest (26.04), independent of this PR: ubuntu-latest moved to 26.04: Linux CUDA release build and Linux arm64 ISA audit fail #132. It blocks publishing v0.8.0.

Hosted evidence (head defccfb)

• Release provenance: pass.
• Native validation: 60/64 jobs pass, including every v0.8.0 row except Linux arm64: Android ISA, Android Vulkan shaders, emulated KleidiAI dispatch, Windows arm64 KleidiAI and Vulkan, Windows x64 runtime closure, MSVC mtmd link, Linux x64 artifact contract, wrapper contract, macOS exit teardown (release and ASan).
• Failing: linux-artifact-contract (arm64) on the pinned v0.7.0, v0.7.0 and v0.8.0 rows, plus the wrapper-validation aggregate. The Linux arm64 CPU ISA audit rejects four SVE instructions outside any sized function in the baseline libggml-cpu.so. The job now runs on ubuntu-26.04 / aarch64 GCC 15.2; the last green main run was on 24.04. It fails on the pinned row too, so this PR did not cause it: ubuntu-latest moved to 26.04: Linux CUDA release build and Linux arm64 ISA audit fail #132.

[leehack]

Independent audit at exact head 16f8a8d against base df1efa1: no P1 or P2 findings, and 64/64 checks pass, including every v0.8.0 row. Evidence is in llamadart-evidence/native-pr-131/.

Corrections to this PR's claims, none affecting code:

• 89 is the unchanged count of required wrapper exports, not the number of new exports.
• Some upstream struct layouts change: mtmd_helper_init_opt grows from 24 to 32 bytes, and mtmd_gen_* change. The core repin must account for both.
• Patch 0003 still ships in v0.8.0, so removing it is a follow-up.

The v0.8.0 CUDA, HIP and Apple release rows have no proof yet. The release fails closed: package-and-release needs every build. Merging with maintainer approval.