Repository navigation
OMP NInfer v0.15.0 — stock OMP 18.14.0 with unchanged GPU lanes
OMP NInfer v0.15.0 — stock OMP 18.14.0 with unchanged GPU lanes
Staged client-only release, not published. Stock Oh My Pi 18.14.0 replaces 18.12.6;
the RTX 5090 container and native Windows RTX 4090/RTX 3090 runtime components, model bytes,
serving profiles and limits remain unchanged from v0.14.0. This release also includes the
product tooling and documentation changes below; it does not rebuild an engine or switch a
native lane to DFlash2.
The published product remains v0.14.0; use its
immutable installation guide
until publication. The prepared guide passed fresh acceptance on
candidate af62e7662fdc7195f7c1c88f2278b15f63130eba; it is not an installation recommendation
before publication. Client/platform, documented-route, public-install and restoration evidence
was re-executed, not inherited. The changed-scope product council is closed by disposition;
the unchanged runtime council retains its original v0.14.0 subject.
Product publication remains founder-owned. Owner-operated, exact-profile 0.x release; no SLA.
Manifest · Qualification ·
Compatibility
Oh My Pi (coding-agent client) connects through OMP NInfer (qualified appliance and release layer)
to NInfer (inference engine) serving Qwen3.8 27B on one user-owned GPU.
Composition — all three runtime lanes carried from v0.14.0
| Component | Exact identity | Serving boundary |
|---|---|---|
| Oh My Pi | Unmodified upstream v18.14.0, source 31147f9d1a76608549821c3759f8487b4d66c72d, tree 9a63b5bfbb5971962c44fca660744b0d75dcd2c2 |
New stock client; fresh three-platform and five-route acceptance passed; macOS remains preview |
| RTX 5090 container | v0.7.0-qwen38-5090-beta.3; server 55f0b95a, archive b8f315bc, image 760fb6be on base d4377ed0 |
Unchanged BF16 KV, DFlash2 K7, two active requests, vision; kept qwen38-5090-v0.13.0 profile |
| RTX 4090 native Windows | v0.7.0-qwen38-4090-beta.3; package 87038cfb, server 5caa7780, public config b5cfbf50 |
Unchanged INT8 KV, MTP3, one active request, Group64 A8 off |
| RTX 3090 native Windows | v0.7.0-qwen38-3090-beta.3; package a7637f66, server fedce425, public config d08d262b |
Unchanged INT8 KV, MTP3, one active request, Group64 A8 off, 350 W serving cap |
All three runtime lanes retain source 2c6f423bd817b6534158553e1b34ab4959c43165 on upstream
64d10fbf9f8e6bfc79ab0570b94674c5b9997929. The model stays the published version 3 artifact
81f924d440c27261d820c19a9f8d45794c5aee410f8a68bd358133fa8c0375da
(20,437,521,664 bytes, revision 1cbd84e7221e51186bd7f093a149912d2489625b). Native packages
retain CUDA 13.3 (sm_89/sm_86); RTX 5090 retains CUDA 13.1 (sm_120a). Runtime receipts retain
their original subjects and dates; they are not new performance measurements or acceptance of
the changed stock client.
Stock Oh My Pi 18.14.0
The new client replaces 18.12.6 without a product fork, local client rebuild, compatibility shim,
or model-ID rename. Full upstream pins are in
client-components.json.
| Platform | Upstream raw asset | Bytes | SHA-256 |
|---|---|---|---|
| macOS arm64 (preview) | omp-darwin-arm64 | 214,816,912 | 608843959ab34295d52460dd09dd7d548b6c903c08a83c67954c2d4c36fd1f5a |
| Windows x64 | omp-windows-x64.exe | 242,797,056 | 620a5935cdb6a9ed52d8b0c93892d6b5bb84155244a98f3be82af0455177bf29 |
| Linux x64 (WSL2 route) | omp-linux-x64 | 276,244,656 | e7dd2b606c4ebe9d1a7bcfad1b8820db968cd6b05409fed1c6bc8db0b6a95a8f |
The host-free client-upgrade review
records the source predicates, all three downloaded asset identities, and two seven-check macOS
stub rehearsals. Those rehearsals demonstrate client semantics, not GPU inference, editing quality,
real NInfer checkpoint survival, or acceptance of this prepared guide.
- Bundled-runtime security: upstream release binaries are rebuilt with Bun 1.4.4. The
18.13.0 release records the fixes for
CVE-2026-91403 (compromised eMudhra intermediate certificates) and CVE-2026-111497 (Brotli
decoder out-of-bounds write); 18.14.0 retains the build pin. Changing YAML or a version label
does not update the old executable. No independent reproducible client rebuild or extraction
of an embedded Bun runtime version is claimed. - Qwen 3.8 replace-edit asymmetry: stock family classification recognizes
qwen3.8-27b, so
native RTX 4090 and theninfer-heavy/qwen3.8-27bselector get thereplacedefault. The opaque
q38-ninferalias used by RTX 3090 and RTX 5090 remainshashline. This is source-reviewed,
not an exercised editing benchmark. Keep exact IDs; do not rename a server or force an edit
policy to trigger family detection. Explicit model variants and strict edit-mode controls remain
owner choices; a settings-derivedhashlinealone is not necessarily a pin. - Exact model selection: retain
ninfer-beta/q38-ninfer,ninfer-native-3090/q38-ninfer,
ninfer-native-4090/qwen3.8-27b, andninfer-heavy/qwen3.8-27b:medium. The suffix selects effort;
mistyped provider-qualified IDs now fail rather than fuzzy-matching a sibling. - Responses and compaction: retain per-model
statefulResponses: true,
includeEncryptedReasoning: false, andsupportsReasoningSummary: false, and clear inherited
PI_OPENAI_STATEFUL. Default compaction for a 131,072-token window triggers strictly above
111,412 tokens, including prompt/tool overhead. The client rework bounds retained history,
ignores stale usage, rejects non-reducing automatic candidates, preserves a completed main
summary if an optional short summary fails (cancellation still throws), and carries split-turn
extra context. These routes do not opt into native remote compaction. Compaction cut points and
prompt bytes can change; production compaction quality and prefix-reuse equivalence are not inferred. - Loop guard: configured
model.loopGuardsettings now reach custom Responses models. Leave
the stock guard enabled; exact suffix-cycle detection and model-gated semantic heuristics are
distinct. Long-loop cutoff behavior and false positives were not measured by the seven-check rehearsal.
Product tooling and documentation since v0.14.0
These are product-side changes, not new runtime bytes or GPU qualification results.
- Native screen arms:
scripts/native_screen_arm.pycaptures
resumable, bounded scratch-install arms with isolated lifecycle state, empty-cache fresh starts,
measured power, sealed shards and a completearm.json. Production runs precede candidate runs;
partial captures are not verdict inputs. Owner-controlled restoration remains a separate obligation. - Quality, fit and parity binding:
scripts/bind_native_variant.py
andscripts/verify_release.pybind and validate native package,
server, configuration and model identities before writing a replacement. A changed runtime needs
a powered quality verdict, or 84/84 actual retained-output-byte comparisons for an unchanged-shape
MTP rebuild. DFlash2 additionally needs a shipping-config 131K fit receipt and lane headroom.
EXP-085's statistical rule is unchanged. The three carried runtime identities do not become new
runtimes merely because these future-qualification tools ship. - Guide-derived client defaults: measurement probes and native campaign launchers derive their
omitted model selectors from the existing guide fragments instead of maintaining another alias.
Explicit selectors remain usable by standalone probes; free-form receipt lane labels do not select
a model. Client acceptance follows the selected route, including RTX 3090'sq38-ninfer, and the
campaign quality child requires the launcher's explicit selector. - Complete WSL2 instructions and safe reruns: the guide includes the Linux client descriptor,
checksum/version checks, local key, loopback route, fail-closed check and restoration direction.
Key/config reruns preserve existing matching files and refuse differing files; Windows native
failures stop the guide. Provider blocks load the key and clearPI_OPENAI_STATEFULbefore a
merged-config refusal. This adds no native Linux OS or Linux vision qualification. - Acceptance and release binding: corrected RTX 5090 attempts need disclosure before effects;
route-runner tests use an ephemeral tunnel port. Composer/verifier checks bind restoration and
route status, G4 ledger agreement, runtime-base receipts and native dependency SBOMs. These checks
do not manufacture acceptance or close every predecessor fault-injection follow-up. - Read-only upstream watch: the watch receipt
records deltas from the shipped engine and client baselines and separately watches
Neroued/ninfer'sfeat/qwen4-27b-opsbranch. Watching a branch does not adopt its code, change
the served model or qualify Qwen 4. - Windows CI coverage: P15 program and lane pushes run the product's CI jobs, including native
Windows portable Python contracts and PowerShell parsing, alongside Ubuntu/macOS contracts and
host-free acceptance dry-runs. GPU-free checks do not prove native CUDA compilation or live serving.
P14 experiment ledger — historical results, not v0.15.0 acceptance
| Entry | Recorded result | Release boundary |
|---|---|---|
| EXP-098 — native DFlash2 gates and fit | On the v0.13.1 control lineage, RTX 3090 passed the 108-role speed gate at ×1.29968 (minimum 0.97987) and a separate 131K fit/true-cold durability proof at 21,968 MiB peak with 768 MiB headroom. RTX 4090's ×1.26362 aggregate missed the per-case floor on log-024@0 (0.945889, reproduced at 0.9451). |
These are historical experimental subjects, not the carried beta.3 MTP3 profile or a new shipping DFlash2 package. Both native lanes remain MTP3. |
| EXP-099 — native small-T tensor-core decode | C1 decode improved +39.8% on RTX 3090 and +8.3% on RTX 4090, but 55/84 and 58/84 role outputs diverged against a maximum of eight. RTX 3090 also failed the 150-pair leak screen (4 baseline-only vs 12 candidate-only; p=0.0384). | Rejected candidate; no small-T tensor-core decode gain ships here. |
EXP-100 — native attention and Group64 A8 prefill |
Byte-identical attention prefill specializations shipped in v0.14.0. The A8 candidate failed the RTX 4090 200-pair leak screen (five ON-only changes, p=1/32); RTX 3090 passed (four ON-only, p=1/16) but the founder kept it off. | The passing RTX 3090 verdict is retained, not rewritten as failure. Group64 A8 stays off on both carried native lanes. |
Unchanged runtime, state and upgrade boundaries
RTX 4090 keeps qwen38-4090-native-v0.7.0-beta.3, package
87038cfb749e00c3af4e590ddcde03f1e7628c8a8c8793b7137f770213865437, public configuration
b5cfbf50d264831832fbe69458b56629a65b0d5b5295ec8bbbd4ea2d73002c38, two extra device-state
slots and 640 MiB admission headroom. RTX 3090 keeps qwen38-3090-native-v0.7.0-beta.3, package
a7637f6672d235e5abb4e54dbdb0689e43a96385b6200c647196aae66b7f42f3, public configuration
d08d262bfe939ed9aaa6f7053ab8a8304fb7e3fef50dad3cb9329e28fb4ece15, four extra slots,
768 MiB admission headroom and its 350 W serving cap. Both remain 131,072-token INT8 MTP3,
one active request and a 30 s pending timeout. Their beta.3 qualification lineages have no
predecessor: rollback is not_applicable, zero directions, not an inherited beta.2 upgrade result.
The v0.13.1-to-v0.14.0 RTX 3090 stop-before-upgrade prerequisite remains relevant only to owners
still making that historical 300 W-to-350 W transition; this client-only release changes neither cap.
No installer fix from a later engine source is retroactively present in these packages.
RTX 5090 keeps qwen38-5090-v0.13.0, configuration
751d22dc405595c17ef129b9b57d3582908567e47e6aa39507c02b9b9da82a47, BF16 KV, DFlash2 K7,
two active requests, and the exact image
ghcr.io/alphastorm/ninfer-runtime@sha256:760fb6bea925eb26d4fa53ab87a126494b69ba7bcfed515d23dee7afa65cdb2a.
The v0.14.0 first-build reproducibility exception remains attached to those same bytes; this
release does not establish a reproducible rebuild. See the
v0.14.0 qualification for the carried runtime observations.
There is no new runtime/profile/store migration in this client-only release. Native stores remain
session-checkpoints-v3 under each immutable component instance; the container retains
/checkpoints-v013. Keep historical roots untouched; do not import generations across different
runtime/model/profile identities. No new rollback direction or universal warm reuse is inferred.
Strict checkpoint origin authentication remains mandatory. Within one API-key principal,
prompt_cache_key names a session capability; use unique high-entropy keys and do not share the
principal among mutually untrusted clients. Saves remain best effort. Windows NTFS
ntfs-write-through is not POSIX fsync or a power-loss guarantee.
Fresh client and route acceptance
Fresh acceptance passed on frozen candidate af62e7662fdc7195f7c1c88f2278b15f63130eba on
2026-10-30. Both initial windows passed and restored; no corrected attempt, canonical promotion
or production upgrade was performed. Runtime qualification carries unchanged, but these changed-client
observations were executed fresh. The receipts are owner-operated tester-equivalent evidence,
not independent external-user outcomes.
| Acceptance evidence | Observed result |
|---|---|
| Exact stock OMP 18.14.0 platform and live tool/continuation/restart/resume/fail-closed checks | All three freshly downloaded stock binaries matched their published asset hashes and omp/18.14.0. Typed read-tool/result, exact continuation and unavailable-route fail-closed checks passed on macOS arm64 (preview), Windows x64 and Linux x64 under WSL2. The macOS documented route also returned the exact nonce after server restart and client resume. Image input passed on macOS/Windows against RTX 5090 only. |
| Final-candidate guide execution: five routes/31 composer-addressed blocks plus WSL2 checks | All five routes passed: RTX 5090 host 2/2, macOS 10/10, Windows 5/5, native RTX 4090 7/7 and RTX 3090 7/7. Pre-publication candidate-clone substitutions are explicit. The WSL2 Linux client additionally passed documented URL/hash installation, typed tools, continuation and runtime-outage checks; this is not native Linux OS or Linux vision qualification. Subsequent guide edits are prose-only; executable blocks are unchanged. |
| RTX 4090 and RTX 3090 public-install identities and observed install outcomes | Both fresh first-lineage documented installs passed with the new stock Windows client after preserving the canonical roots and staging. Separate RTX 4090 and RTX 3090 public-URL checks verified all manifest asset hashes and served identities; both installers returned already_installed for the unchanged beta.3 packages (87038cfb / a7637f66), changed no lifecycle pointers and requested no runtime start. Both lineages have no predecessor: rollback not_applicable, zero directions. |
| Independent container/native restoration audits and observed hold durations | Independent restoration passed. RTX 5090 restored the same digest-pinned 760fb6be image, healthy and unless-stopped, checkpoint mount, checkout, key, directories and Windows markers/tasks; the route container and hold were removed after 403.884 s. Both native restoration/final audits passed, including state/ACL/task/power/holds/client files/directories/staging, with cross-lane census unchanged; driver windows were 371.326 s (RTX 4090) and 394.389 s (RTX 3090). Canonical beta.3 tasks returned Ready/stopped at owner baselines 450/370 W, without a process, listener or lease. |
The composed acceptance binds these fresh
observations separately from carried runtime qualification. Root compatibility authority, profiles
and launchers were lane-pinned to this candidate. The bound review ledger
records the changed-scope product council closed by disposition and the unchanged v0.14.0
fork-council entry. No in-scope residual P0/P1 remains after the recorded lead controls.
Final manifest evidence pins, strict ready verification and founder-owned publication are
separate controls; acceptance and G4 approval grant no publication authority.
Support boundaries and next work
macOS remains preview; Linux client evidence is WSL2-scoped. Native lanes serve sequentially on
port 18082. Native vision or DFlash2 serving, concurrent native serving, multi-GPU inference,
multi-tenancy, silent cloud fallback and independent external-user outcomes are not qualified.
The carried runtime council retains its original v0.14.0 subject, reproducibility exception,
P2 deferrals and missing Mythos transport result; it is not the new product-tooling review.
All four seats completed the new product council; its P2/P3 residuals are deferred in
#117, including the native-screen
binding prerequisites before v0.15.1 can bind the RTX 3090 DFlash2 screen.
The attempted N15/F2 fork-council residuals in
#116 concern deferred, unshipped work;
that council does not approve any new runtime in this client-only release.
The alphastorm/ninfer#48/#49 engine/installer fixes are not present in any of the three carried
runtime components. RTX 4090 DFlash2 K6/K5 exploration and its failing floor case remain deferred.
Next: native RTX 3090 DFlash2 is deferred to v0.15.1 after a found runtime defect:
fork-after-resume fails with Native could not release an unreferenced checkpoint owner.
The fix and shipping-package qualification remain required; no delivery date or unmeasured
performance, quality or durability outcome is claimed.