OMP NInfer v0.12.0 — Native RTX 3090 and RTX 4090 on one two-GPU Windows host

OMP NInfer v0.12.0 — Native RTX 3090 and RTX 4090 on one two-GPU Windows host

Owner-operated, exact-profile 0.x release; no SLA. Native Windows RTX 3090 and
RTX 4090 lanes return on one Windows host with both GPUs installed. Each lane binds
its own card by GPU UUID. Only one native lane runs at a time: both use authenticated
loopback port 18082. The RTX 4090 has no supported rollback target on this topology.
This is not multi-GPU inference, concurrent native serving or an automatic lane switch.

The RTX 5090 Windows 11 Docker Desktop/WSL2 runtime, its model and serving profile,
and unmodified upstream OMP 18.8.7 are unchanged from v0.11.0. The container route
uses authenticated local loopback or a manual macOS SSH tunnel. All five documented
routes, three stock clients and both native public installs passed fresh acceptance.

Manifest · Qualification ·
Quickstart · Security model ·
Known limitations

What changed

• Native lanes return with explicit GPU ownership. Lifecycle scripts resolve the
  installed release's UUID to the current device. GPU-owner, acceptance and VRAM
  queries and power changes address only that card. A legacy record without a UUID
  fails closed when more than one GPU is visible; hiding a card with
  CUDA_VISIBLE_DEVICES is unsupported.
• The RTX 4090 starts a new supported lineage on this host. Its rollback result is
  not_applicable, zero directions, first release of the lineage. v0.6.10 and earlier
  RTX 4090 packages use single-visible-GPU lifecycle operations and are unsupported
  here. Acceptance uses --first-release, not a fabricated predecessor or rollback pass.
  The RTX 3090 retains its actual two-direction qualification against v0.6.2-beta.1.
• The separate native model binding returns. v0.11.0 removed it when deferring
  both native lanes. Restoring v0.10.0's exact native artifact avoids binding these
  packages to the different RTX 5090 model; neither model is changed.
• Fresh product proof, unchanged client. Native qualification used OMP 18.0.9.
  Product acceptance separately exercised the pinned stock OMP 18.8.7 binaries;
  no previous product-route or platform-client proof is carried forward.

Exact component identities

Both published native components use source
cad1976764fded5fbe96930c36bed0570b3d3de7 and CUDA 13.3 lifecycle packages.

Native lane	Component tag	Package SHA-256 / bytes
RTX 3090	v0.6.3-qwen38-3090-beta.1	74120de0e6b1037e3c994e21b299f5cf4b98d9a3afa4e137d1c2ce8b57bfc297 / 595,697,859
RTX 4090	v0.6.11-qwen38-4090-beta.1	526870444c71add22c3dff8beb0c767aab2cbbb98870ec060398e6b7243de3a2 / 574,770,447

The native model is the 18,210,531,328-byte artifact at revision
18dfc887423fa5aabf3cb56fac41490e462b3fab, SHA-256
eec39564993d6e9c7d5e383382a760f093465c9d163ec9a1bd6b80199514bf3e.
The manifest binds each package, source archive, checksum set and build receipt.

The unchanged RTX 5090 component is v0.6.16-qwen38-5090-beta.1, source 1302d639,
server 548fe239, published image 6a02feba, model 0634abb0 and deployment profile
qwen38-5090-v0.11.0 / configuration 91a35670. Stock OMP 18.8.7 remains pinned to
source f261ed9faf16b61880b544f599876bface4ded0d and the same three upstream binaries.
Full identities and serving limits are in the manifest and compatibility matrix.

Qualification and acceptance

The public RTX 3090 and
RTX 4090 receipts bind the measured packages and profiles.
These measurements are not a performance improvement claim or universal GPU figures.

Native qualification workload	RTX 3090	RTX 4090
Exact 130,048-token retrieval	220.5 s	91.0 s
C1 decode	102.4 tok/s	158.7 tok/s

The RTX 3090's explicit 42-token checkpoint and previously unpublished 45-token
session survived a graceful managed restart. Checkpointing remains best effort;
these short controls do not prove globally loss-free shutdown. The RTX 4090's
no-target rollback result remains visible in both qualification and acceptance.

All five documented routes, 31 blocks and
three stock OMP 18.8.7 clients passed on frozen candidate
7713ceb956eee79f4c3e232ca020d385baff89be. Native routes ran sequentially. Both
public URL installs passed with the verified model path derived from the active
installed release record. Composed acceptance
binds the public-install receipts and behavior/restoration proof.

The native windows restored their measured state, pointers, release counts, tasks,
ACLs/secrets, holds and UUID-selected power. The RTX 5090 window restored its retained
incumbent image, mounts, health, key and checkout baseline in 389.939 s; production
was not upgraded. Platform, acceptance and manifest pins bind the final evidence
bytes without advancing the executable subject or changing any executed block.

Support boundaries

• Only the documented exact profiles are qualified; no smaller-host native profile
  or unattended RTX 3090 role activation is claimed.
• RTX 4090 rollback is unavailable on this two-GPU host until a real supported
  predecessor exists. RTX 3090's two-direction result does not transfer to it.
• macOS remains preview. The Linux client ran under WSL2, not a separately qualified
  native Linux OS. Stock macOS/Windows image-input checks do not establish native-lane
  vision support, structured JSON-schema output or a new client feature.
• No automatic container restart, multi-tenancy, universal warm reuse, independent
  external-user outcome, support response-time guarantee or SLA is claimed.
• Acceptance restores the recorded baselines; it does not prove restoration of
  originally absent mutable directories. That successor correction remains separate
  and must not merge before this candidate is published.

The candidate is technically ready, not yet a published product. Only the owner may
tag/publish v0.12.0; normal install blocks require that ready product tag. No release
date is assigned here, and no product publication is authorized by these notes.