NInfer RTX 5090 durable runtime source (v0.4.5)

Exact source archive for the v0.4.5 runtime binaries (dde8d07: manifest origin authentication).

NInfer 0.4.5 Qwen3.8 RTX 5090 runtime beta

Digest-addressed linux/amd64 CUDA 13.1 runtime image built from v0.4.5-qwen38-5090-beta.1. No mutable consumer tag is published; use image_reference from the receipt.

NInfer RTX 5090 durable runtime binaries (v0.4.5)

Checkpoint manifest origin authentication (closes #32): every save writes manifest.mac - HMAC-SHA256 over the exact manifest bytes, keyed by material derived from the bearer key and held outside the checkpoint root. Loads verify origin before trusting manifest content; transient tag faults preserve current for retry; --session-checkpoint-require-origin-auth is the strict NAS/S3 import posture with a reversible refusal and a 32-character bearer floor. Additive and rollback-safe: prior binaries read the same store. Council CRS-origin-auth closed 7/7; durable arc receipt in omp-ninfer docs/measurements/2026-08-31-origin-auth-probe.json. Supersedes v0.4.4-qwen38-5090-beta.1.

NInfer RTX 5090 durable runtime source (v0.4.4)

Exact source archive for the v0.4.4 runtime binaries (51fd30e: buffered checkpoint export + covers() redundant-save skip + idle debounce + lazy-restore session repair). Supersedes source.1 of v0.4.3.

NInfer 0.4.4 Qwen3.8 RTX 5090 runtime beta

Digest-addressed linux/amd64 CUDA 13.1 runtime image built from v0.4.4-qwen38-5090-beta.1. No mutable consumer tag is published; use image_reference from the receipt.

NInfer RTX 5090 durable runtime binaries (v0.4.4)

Checkpoint export decoupling: exporter writes flow through a bounded in-memory queue drained off the engine execution lock (fail-before-publication preserved); automatic saves debounce to sustained-idle, skip frontiers already catalogued, and never queue foreground requests behind an elective save; lazy restore repairs partially resident sessions. Measured on the identical package: warm edit 15.26 s -> 0.91 s, explicit save 31.6 s -> 13.8 s, four fanout branches 0.90-1.01 s with auto-save at defaults. Council CR-20260831-ckptdecouple closed; receipts in omp-ninfer docs/measurements/2026-08-31-fanout-probe-v044c.json. Supersedes v0.4.3-qwen38-5090-beta.2.

NInfer RTX 5090 durable runtime source (v0.4.3)

Exact source archive for the v0.4.3 runtime binaries (82a1a22; fanout anchors + checkpoint import integrity + session isolation + guard closures). Supersedes source.1.

NInfer 0.4.3 Qwen3.8 RTX 5090 runtime beta

Digest-addressed linux/amd64 CUDA 13.1 runtime image built from v0.4.3-qwen38-5090-beta.2. No mutable consumer tag is published; use image_reference from the receipt.

NInfer RTX 5090 durable runtime binaries (v0.4.3)

Fanout train: same-lane branch fanout anchors on /v1/responses (sibling branches of one previous_response_id reuse the base prefill), checkpoint-import TOCTOU closure with streamed re-hash + corrupt-generation quarantine, symlink-refusing export writes, constant-time credential equality, compute-to-transfer export fence, id-correlated refusal diagnostics, media-enum rejection and streaming UTF-8 repair picks, plus the session-isolation set (private continuations never cross sessions) ported from the qualified 4090-durable train after the agent-protocol smoke exposed the latent 5090 cross-session private reuse.

Source: 82a1a22 (upstream base 6e8b2e2). Built in pinned runpod/pytorch:1.1.0-cu1300-torch291-ubuntu2404 on the owner appliance. 9/9 ctest green; agent-protocol + serve-contract smokes rc=0; source_dirty=false. Council CR-20260831-fanout43 closed with dispositions. Supersedes v0.4.3-qwen38-5090-beta.1 (immutable; bound a pre-isolation binary).

Measured (67.7K-token base, 4 branches): total fanout wall 148.7 s -> 47.9 s; device-resident branch ttft 0.40 s (was 26.4 s); post-restart resume green. Sibling-branch KV cloning ceiling tracked in #34.

sha256: 04d501a2760b6846fb52f039721c77fb7a175620aee7fc4208ee04a57a383e0c

NInfer RTX 5090 durable runtime source (v0.4.1)

Exact source archive for the v0.4.1 runtime binaries (f4fb315; quota transition + skip reasons + post-publish reclaim acknowledgment).