OMP NInfer v0.4.0 — durable session state on every lane

Session state survives the process now — on every lane. The RTX 5090 container gains transactional durable session checkpoints: an automatic 7.95 GB checkpoint at a 109,725-token frontier restored 109,589 tokens hot across a docker restart on a rotated server instance — 0.778 s to first token serve-side, versus 47.92 s to rebuild the session cold. Exact retrieval at 130,448 tokens; decode 144.80 tok/s on the agent-shaped gate (152.2 at 83.3% MTP acceptance on retrieval); the pinned OMP 18.0.9 client accepted the live status and checkpoint documents verbatim. Durability now ships on all three qualified lanes — the native Windows 4090/3090 DirectStorage lanes since v0.3.x, the 5090 container as of this release.

New runtime image ghcr.io/alphastorm/ninfer-runtime@sha256:8de5efdf… (source 1ceaeebd), reviewed by a cross-family council with every mitigation landed before freeze — receipts ship with the component releases (binaries · image receipt · source). The previous image digest remains published as the rollback target. Also new: examples/fleet/ for running three model-bound agents across the lanes.

Start here: Quickstart. 4090/3090 lanes unchanged from v0.3.2.

OMP NInfer v0.3.2 — corrective 4090 qualification summary

Corrective release. The v0.3.1 RTX 4090 qualification summary carried its v0.2 template's limitations — "MTP0 qualified", "MTP3 performance not claimed" — in direct contradiction of the MTP3 receipts it fronts, plus a stale beta classification. v0.3.2 corrects that summary; no component bytes, receipts, or measured numbers changed. All install lanes are identical to v0.3.1, which remains immutable with the defect on record.

Start here: Quickstart.

OMP NInfer v0.3.1 — qualified RTX 4090 MTP3 lane

The RTX 4090 lane now ships its qualified MTP3 speculative profile. Identical released engine binary and model bytes — only the speculative configuration changed (a +4-byte deterministic package) — promoted by the recorded two-arm MTP0-versus-MTP3 comparison: complete Golden-equivalent wall time 5.663 s → 4.698 s (+17.04%), decode 93.2–97.7 tok/s (vs the 52.330 tok/s MTP0 baseline), 107,851-token restored continuation after process replacement, every deterministic, persistence, and protocol gate green on both arms. An exploratory depth sweep measured draft-3 > draft-4 > draft-5 on the fixed decode workload — the first datapoint for the roadmap's MTP depth-and-corpus ablation.

Start here: Quickstart. RTX 5090 and 3090 lanes are unchanged from v0.3.0.

Component release: ninfer@v0.3.1-qwen38-4090-mtp3.2 (immutable, SHA256SUMS-bound). Every number above is receipt-bound in the attached evidence set; qualification tooling defects found during the campaign are disclosed in the qualification sidecar.

OMP NInfer v0.3.0 — first public release

The first public release. If you use Oh My Pi and own an RTX 5090, 4090, or 3090, OMP NInfer keeps Qwen3.8 27B and your session's continuation state on your GPU — as an explicit, durable primitive, not a lucky prefix-cache hit. (A live server's prefix cache already covers the append-only case; what ships here is transactional lineage plus native-lane checkpoints that survive process restarts.) Start here: Quickstart.

Three qualified lanes, one manifest

Lane	Decode	Highlights
RTX 5090 · container	240.30 tok/s	Freshly requalified on the identical published bytes; 2,199.41 tok/s prefill at 130,048 tokens (exact retrieval); warm follow-up 0.191 s (retained state) vs 36.651 s cold (fresh process, no cache) at an 89,022-token session — qualification-bound
RTX 4090 · native Windows	52.330 tok/s	102,075-token restored continuation after process replacement; DirectStorage-backed durable checkpoints
RTX 3090 · native Windows	90.17 tok/s	New in v0.3.0: full public install authority — 15/15 protocol, exact 64,512-token retrieval, durable restart, bidirectional rollback, 299.8 W peak at a 300 W cap

Durable session checkpoints ship on both native Windows lanes (DirectStorage) — a process restart resumes from restored state instead of re-prefilling. The RTX 5090 container keeps live-process warm continuation with OMP transcript replay as its recovery path; per-lane semantics are explicit in the receipts.

Verify everything

git clone --branch v0.3.0 --depth 1 https://github.com/alphastorm/omp-ninfer.git
cd omp-ninfer
python3 scripts/verify_release.py --require-ready

Model by SHA-256, runtime image by OCI digest with SPDX SBOM, clients by checksum, native packages by SHA-256 — one manifest binds them all, and the attached qualification receipts carry every number quoted anywhere. Benchmarks with charts · Architecture · Security model

Support boundary

0.x series: the latest published release and its exact manifest/profile receive fixes. One owner-operated machine per lane; one active request per profile; loopback-only, bearer-authenticated, fail-closed — an unreachable GPU errors, it never silently answers from a cloud model. Community project; not affiliated with or endorsed by Oh My Pi, Qwen, or NVIDIA.

OMP NInfer v0.2.0 beta 1

OMP NInfer v0.2.0-beta.1

Invited-tester prerelease. Not stable or GA. No production route is activated.

What is qualified

• Native OMP 18.0.9 clients for macOS arm64, Windows x64, and Linux x64, with exact public source at alphastorm/oh-my-pi.
• Manifest-driven doctor, plan, install, status, quick benchmark, checkpoint, rollback, and support-bundle adapters.
• Primary RTX 5090 runtime: digest-pinned BF16-KV/MTP3 image, 131,072-token context ceiling, 130,048-token exact retrieval, and 235.02 tok/s over 2,048 decode tokens.
• Native Windows RTX 4090 beta: exact MTP0 package, 15/15 protocol, 102,060 → 102,075 process-restart continuation, 1,410.691 prefill tok/s, 52.330 decode tok/s, exact source-controlled OMP Golden-equivalent, and full lifecycle/restoration evidence.

RTX 3090 preview

The tagged RTX 3090 package is public and review-closed but not installable or beta-qualified. Current sm_86 response-store, session-checkpoint, and HTTP contract tests passed 3/3 on Windows and an ephemeral Community RTX 3090. The following current-package gates remain not_run:

• live authenticated model protocol;
• 64,512-token retrieval;
• process-restart continuation;
• bounded C1 performance;
• fresh exact Windows RTX 3090 install/security/bidirectional rollback; and
• exact OMP 18.0.9 acceptance.

Community live-model attempts returned provider cudaErrorUnknown; all owned pods were deleted and no passing claim is inferred. Historical package measurements are isolated and do not qualify the preview package.

Evidence and safety

• Manifest
• Product qualification
• Compatibility authority
• Human-readable matrix
• Quickstart
• Security model

The historical private Golden corpus was unavailable and was not reused, read, copied, hashed, or transmitted. Its RTX 4090 replacement is source-controlled and proves one typed tool call, exact primitive arguments, linked tool-result continuation, and an exact visible final answer.

Process-restart Responses continuation is qualified only for the native RTX 4090 variant. The primary RTX 5090 image uses restart policy no; OMP transcript replay remains its recovery path. Response DELETE is logical object deletion, not secure erasure of context required by surviving descendant continuation.

Structured JSON-schema output, multi-GPU, multi-tenant, priority, preemption, silent cloud fallback, and universal performance claims are excluded.

OMP NInfer v0.1.0-beta.1

OMP NInfer v0.1.0-beta.1

Invited-tester prerelease for one trusted RTX 5090 owner. The ready primary route is native Windows OMP with Docker Desktop WSL2; managed macOS SSH and native Linux clients remain preview profiles.

Install

Follow the tagged early-access quickstart. The ready contract must pass before setup:

python3 scripts/verify_release.py --require-ready

Primary Windows component:

• release: omp-18.0.7-cross-platform-preview-5
• archive SHA-256: 4d94ffd7761558691cc5381d56069ea9d3bd2b081d37ce5afbaf2c05ef9af5a7
• binary SHA-256: 0b59681010427a77aa39f7929576e270e91f8e8cc563a9f32f865090fc9ec892

NInfer runtime:

• image: ghcr.io/alphastorm/ninfer@sha256:a19e00c5d7993d7c12341974bdc7e3d13e61d9deaf39e8ce00cf94abd6137712
• server binary SHA-256: af557e9b146c68b051bd29e9e7a7d172e908f4b12df4e6d3bd83c54ce58b2399
• model SHA-256: eec39564993d6e9c7d5e383382a760f093465c9d163ec9a1bd6b80199514bf3e

Qualification

The owner-operated tester-equivalent clean Windows path downloaded only the public client and compatibility authority, then passed exact install/version identities, Docker/GPU doctor, authenticated status, file tools, Vision, stateful exit/resume, runtime-outage fail closed with no cloud response, and exact production-runtime restoration.

Runtime qualification separately covers 131,072-token context, OpenAI/Anthropic/Responses schemas, Vision, stateful continuation/forks/delete behavior, prefix reuse, Golden behavior, 209.04 decode tokens/second, and 0.76975 MTP acceptance on the recorded RTX 5090 profile.

Known boundaries

• This is early access, not general availability.
• One active request, one resident model, one trusted owner.
• No automatic container restart or NInfer process-restart continuation claim.
• macOS managed SSH and native Linux clients remain preview profiles.
• RTX 4090, multi-GPU, multi-tenant, and broad hardware claims are deferred.
• Ordinary tester defects will be fixed in the next release; release-integrity or safety defects may require withdrawal.

The product manifest, qualification summary, compatibility authority, acceptance receipts, and changelog at this tag are the release authorities.