feat(abi): explicit device selection through the ONE surface - vllm_model_params.device, ABI v14, --device on server + cli

[localai-bot]

ONE-SURFACE fold ROW 8 (row/DEVICE-KNOB, task #284) — the smallest fold row: an embedder can now ASK for a device instead of inheriting the accelerator-first probe. Mirrors vLLM's DeviceConfig.device names (vllm/config/device.py:13 @ 555967922; explicit assignment is VERBATIM, never substituted, device.py:61-66 — upstream's own --device CLI flag was removed in v0.10, bc8a8ce5e, so the field mirrors the surviving config surface).

Rebased onto main @ 9f772cad (#122 landed): ABI bump is v13 → v14; changelog reads v10..v14 chronological (the #123-review v12-move minor kept); c_header_compile.c composes #122's vllm_complete_tokens reference with the v14 field; allowlist ratchet inherited UNTOUCHED at 8.

What

• W1 ABI: vllm_model_params.device appended (int32, 0=auto / 1=cpu / 2=cuda; 0 MUST be auto for the zero-value contract; cpu-before-cuda follows the v12 vllm_video_model_params.device precedent shifted by the auto slot). VLLM_ABI_VERSION 13 → 14; floor pin >= 14 (the one-surface(parakeet): fold Parakeet ASR onto the C ABI (ROW 1) #121 == lesson). New include/vllm/config/device.h + src/vllm/config/device.cpp (vllm::Device, DeviceFromString, DeviceName).
• W2 plumb: EngineParams::device → SelectQueue(architecture, device); explicit arms route through the pure LoadedEngine::ResolveExplicitDeviceType(device, cuda_platform_registered) so the CPU tier gates the full matrix, including explicit cpu beats a REGISTERED accelerator. Explicit cpu never probes; explicit cuda requires the registered kCUDA platform and THROWS a pinned message otherwise; explicit-arm queue-creation failures PROPAGATE (auto keeps its byte-identical catch-and-fall-back). FromModelDir resolves an explicit device BEFORE any path/config I/O (mirror of DeviceConfig resolving at config time, arg_utils.py:1878) — a device error is never masked by a path error, which is also what makes the capi plumb pinnable with no loadable checkpoint. capi validates 0/1/2 (else VLLM_ERR_INVALID_ARGUMENT); explicit cuda on the CPU-hosted transcription stack is refused, not silently downgraded.
• W3 thin clients: --device auto|cpu|cuda on vllm-cli (pure ABI field consumer) and examples/server (vllm::DeviceFromString → EngineParams.device; --video-device untouched — separate engine, separate checkpoint).
• W5 (feat(abi): MiniMax-H3 video+audio generation through the ONE surface — vllm_video_*, ABI v12, /v1/videos via the library seam, thin-client examples #123 review's 3 MINORS, tests/docs only): c_header_compile.c now actually references the v11 transcription + v12 video entry points + the v14 field (and, post-rebase, rides alongside model(kimi_linear): ONE-SURFACE ROW 7 — fold Kimi-Linear onto the shared paged runner #122's vllm_complete_tokens reference); the v12 changelog block moved to chronological position; the H3 fold fixture flag list gained --keep-quant.

Gates (all green on the rebased head 3f8de74f)

Full CPU build -Werror clean (pre-rebase full tree; rebased head rebuilt + re-run on the gate set) · test_capi 44 cases / 428 asserts (4 new v14 + floor >= 14; #122's v13 cases green alongside) · test_loaded_engine_dense 65 asserts (+3 device cases incl. the pure policy matrix) · test_dlopen 28 · test_openai_api_server 545 (+explicit-cpu serving smoke) · test_minimax_h3_video_fold green (vllm.h touched) · test_model_loader_gguf + test_platform green · check-surface-coverage OK — 8 internal-reachers, ratchet UNTOUCHED at 8, capability table 11 reachable / 2 tracked · scripts/agent-preflight.sh EXIT=0 (record checkers + mutation suites, --staged green, gate-chained pushes) · STATUS ratchet paid by removing a stale back-to-back merge duplicate (279200 → 279150, measured 279111 post-rebase).

Mutation kills (each RED, then reverted; recorded in .agents/state.md)

M1 default device=2 → zero-contract 3 fails · M2 capi 0→kCUDA → zero-contract bogus-path arm RED · M3 range validation dropped → 6 fails · M4 capi drops the device=2 mapping → plumb pin RED · M5 FromModelDir early resolve removed → capi plumb pin + device-before-path test RED · M6 ctor passes kAuto → EngineParams-seam ctor-throw arm RED · M7 SelectQueue explicit-arm catch-and-fallback → same test RED · M8 policy hands explicit-cpu to a registered accelerator → matrix RED · M9 DeviceFromString("cuda")→kCPU → parse test RED · M10 ABI macro left at pre-bump → floor pin RED. Disclosed not-uniquely-killable: the api_server smoke (same seams as M6-M8), c_header_compile (compile-time reference), the two doc-only minors.

Residuals (named)

1. CUDA-build A/B (auto→CUDA vs explicit-cpu→CPU vs explicit-cuda→CUDA on a GPU box) — no dgx in scope; the CPU tier pins that half via the pure matrix.
2. Explicit names for xpu/vulkan/metal/rocm are additive later; auto reaches them today.

🤖 Generated with Claude Code

https://claude.ai/code/session_014fZAcg1WcU8V629k6HWKys

[localai-bot]

Landed on main as bcd9c48 (mudler-authored squash via commit-tree). Fresh mutation review MERGE-CLEAN (zero-contract, no-fallback, never-probes, resolve-before-IO, v14 floor pin all independently re-killed; device.py port verified verbatim at the parity pin; shared model_loader blast radius green incl. the Kimi paged load). Operator landing: merged tree byte-identical to the reviewed head, four row suites 4/4, eleven gates 00000000000. This completes the session's fold arc: ABI v10 (text-only) -> v14 (transcription, video generation, token-level completion, device selection), surface-guard ratchet 12 -> 8, all enforced in agent-preflight.sh + CI. Named residual: CUDA-build A/B of the explicit-device matrix on a GPU box.

[localai-bot]

Post-landing note: bcd9c48's parent absorbed a concurrently-landed main (#135) due to a symbolic-ref race, transiently reverting #135's content; repaired in b44ad33 with full disclosure there. #136's own content was and remains intact.