Skip to content

DeepSeek V4 Flash 0731

Current status and review date

Decision snapshot

  • Product role: former human-approved text Primary; retained as reproducible historical evidence, not a claim about the current live route.
  • Selected or best-qualified configuration: Infernal Invocation r18 B12X with DSpark K5, TP=2, 1,048,576 tokens, eight admitted sequences, and 4,096-token batching.
  • Measured hardware: two NVIDIA RTX PRO 6000 Blackwell Max-Q cards in exclusive TP=2 over PCIe without NVLink.
  • Evidence: DSpark K5 raised median decode from 76.4 to 142.1 tok/s at 4K and from 76.3 to 129.5 tok/s at 32K versus the otherwise matched no-spec control; the profile also passed the retained 1M, functional, quality, concurrency, and real-client gates.
  • Decision: r18 became historical when Qwen3.8 Flash Next was promoted on 2026-08-26. Earlier r15 and r33 profiles remain dated recovery and comparison evidence.
  • Important limitation: native CPU KV offload is a separate 262,144- token WSL2 capacity lane, not part of r18; representative SWE scoring, some routed >300K gates, and automatic rollback remain unproven.
  • Review dates: retained evidence through 2026-08-26; dossier-format review 2026-08-31.

Open the retained container configuration inventory or jump to the decision, known limitations, or dated evidence.

Review narrative

The dated notes below preserve the full reasoning behind each decision. The later measurement sections retain the exact metrics, controls, failures, and promotion boundaries.

2026-08-21–26 — r18 1M qualification, promotion, and supersession

The 2026-08-21 record qualifies and promotes Martin Vit's Infernal Invocation r18 B12X TP=2 DSpark K5 profile at 1,048,576 tokens, maxseq8, and 4,096-token batching. It passed 1,040,063 actual prompt tokens, repeated agentic checks 12/12, structured tools 160/160 plus the complete functional batches, client-shaped reserve probes, short c8 and long c2 capacity, a matched no-spec A/B, and a clean reload gate. After operator authorization, the guarded transaction, authenticated routed functional gate, Mini generation-2 convergence, and real Hermes/Pi/OpenClaw acceptance passed. r18 remained the text llm.primary until the 2026-08-26 Qwen3.8 Flash Next promotion and is now retained historical evidence.

Outcome: the exact r18 K5 profile qualified at the model's declared 1M limit, passed the guarded client-facing transaction, and is now retained as a former Primary rather than presented as current deployment state.

2026-08-16 — r15 393K promotion

The 2026-08-16 public record documents human approval of Martin Vit's Infernal Invocation r15 B12X TP=2 DSpark K5 profile for the text llm.primary route at 393,216 tokens, 4,096-token batching, and eight admitted sequences. It passed 351,118 actual prompt tokens directly, 340,119 through the authenticated router, repeated tools/session/diff/timeout checks 12/12, short c8 and long c2 capacity, and OpenClaw-compatible Anthropic wire calls. The r15 record is now historical. The prior r33 393K profile was the fixed-port managed recovery target while r18 was current.

Outcome: r15 qualified and was promoted at 393,216 tokens, then became historical after r18. Its evidence remains a matched performance and routed acceptance reference.

2026-08-02–11 — native offload, r33 controls, and GPU-only capacity

A derived WSL2 image also qualified native CPU KV offload at a 262,144 served-token ceiling. It passed a cold ladder through 249,573 prompt tokens and the managed shared-memory lifecycle regression. This is retained historical capacity evidence and is not part of the current r18 profile. A digest-pinned r33 target-only control subsequently qualified the same checkpoint at a 131,072-token envelope and reached 119,503 actual prompt tokens. A matched batch-token A/B then reduced profiled activation memory by 34.1% and increased minimum-rank KV allocation by 0.72 GiB while retaining the same functional and 119,503-token capacity gates. The later r33 DSpark arm reported 725,543 GPU KV tokens and passed 238,507-, 339,310-, and 359,900-actual-token direct requests without host offload. Review date: 2026-08-26.

Outcome: these lanes establish bounded WSL2 offload and GPU-only capacity results, but they do not extend the r18 contract or establish current routing.

Earlier and still-open evidence

The earlier SGLang 32K low-reasoning lane remains valid point-in-time evidence, not the current performance recipe. Community 0731 NVFP4 and GGUF conversions remain unqualified. The local maxseq16 lane has bounded c16 and three-tool-burst evidence, but not broad repeated multi-agent quality.

Outcome: retain these as historical controls and research leads; do not promote their unmeasured or incomplete properties into the qualified contract.

Immutable identity

The r16 result pins deepseek-ai/DeepSeek-V4-Flash-0731 revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb, the release revision containing the tested weights. The upstream repository has since advanced to 7872f01b1d1fe23eabc4c98b48bffcef5a386062 for model-card documentation; that later commit was not substituted into the measured recipe.

DeepSeek identifies 0731 as a re-post-trained official release that supersedes Flash Preview while retaining the V4 Flash architecture. Treat it as a distinct behavioral generation. The target model is described as 284B total and 13B active; checkpoint inventories that include the bundled DSpark draft module report approximately 304B. It is MIT licensed with a declared 1,048,576-token maximum context.

The 2026-07-10 nvidia/DeepSeek-V4-Flash-NVFP4 attempt used NVIDIA's Preview conversion, did not retain a reusable checkpoint revision, and remains historical-invalid. It is not a measurement of 0731.

Tested hardware and topology

Two RTX PRO 6000 Blackwell Max-Q cards on Primary Node in exclusive TP=2 over PCIe without NVLink. Every other inference workload was offline. The 192 GB figure is aggregate VRAM, not unified memory.

Engine, quantization, KV, context, and concurrency recipe

For reusable TOML, container field mapping, and standalone Docker reconstruction guidance, open the container configuration inventory.

Former r18 1M profile

The former r18 profile uses image digest sha256:414ec7d0d28358cfd8af0697f330f5c8acbb80e4dc4e5ba69c9fd5b5855ea804, vLLM integration tree f0fa1cefc1865d316c2478525f550e7646addc40, B12X tree 75787c7a7431b3bea414d2ebf5f2b8671b23eb33, W4A8 kernels, FP8 compressed MLA KV, InstantTensor BUFFERED, TP=2/DCP=1, 1,048,576 context, eight engine sequences, and fixed probabilistic DSpark K5. Native KV offload and LMCache are disabled. The matched performance workload is c1; capacity also covers c8 short requests and c2 at 490,861 prompt tokens per request. Router admission is one full-window request.

The native-Linux source recipe required an SM120-to-WSL2 translation: direct NCCL P2P and NCCL cuMem device/host allocation are disabled, shared memory remains enabled, and PyTorch expandable segments are disabled. Model, JIT, and temporary build data use named Docker volumes.

r16 DSpark, offload, and extended-context recipes

r33 target-only, batch, and DSpark recipes

Infernal Invocation r15 recipes

Infernal Invocation r18 recipes

Evidence by measurement class

Each subsection answers a distinct reuse question. Performance numbers remain bounded to the exact model, runtime, topology, context, and concurrency stated for that lane.

r18 — 1M performance and client acceptance

The r18 DSpark K5 profile adds functional, capacity, matched performance, and bounded quality evidence at the model's 1,048,576-token limit. Calibrated retrieval passed through 1,040,063 actual prompt tokens; agentic intelligence/session/tools passed 12/12; an additional structured-tool soak passed 160/160; client-shaped 5,120- and 8,192-output-reserve probes passed; c8 short and c2 at 490,861 prompt tokens per request completed. The engine reported 1,323,176 KV tokens, or 1.26 full windows. Against the matched no-spec control, K5 raised median decode from 76.4 to 142.1 tok/s at 4K and 76.3 to 129.5 at 32K. The operator-authorized live transaction then passed exact routed identity, the complete routed API/tool subset, generation-2 Mini convergence, and real Hermes/Pi/OpenClaw acceptance.

Outcome: qualified at the declared 1M context with matched DSpark gains and complete promotion-era client acceptance; now historical.

r15 — 393K performance and routed acceptance

The r15 DSpark K5 profile adds functional, capacity, matched performance, bounded quality, and routed acceptance evidence. K5 versus the otherwise identical no-spec control measured 150.0 versus 76.4 tok/s median decode at 4K/c1 and 119.245 versus 76.767 at 32K/c1. Direct retrieval passed at 351,118 actual prompt tokens and authenticated routed retrieval at 340,119. Repeated tools, session recall, unified diff, and parallel timeout triage passed 12/12. The engine reported 797,689 KV tokens, or 2.03 full 393,216-token windows. The stock English 390K routed needle still failed closed under the router's conservative byte estimator, and generic thinking-disable did not suppress r15 reasoning. Those are retained limits, not relabelled successes.

Outcome: qualified former 393K Primary with strong matched DSpark gains; the conservative routed needle and reasoning-disable behavior remain explicit limits.

r33 — GPU-only 393K profile

The historical r33 DSpark profile adds functional, capacity, and bounded quality evidence: exact managed routing passed, the direct context ladder reached 359,900 actual prompt tokens, and OpenClaw/Hermes client-path requests passed after safe gateway restarts. It uses FP8 DS-MLA KV with no host offload and reported 725,543 GPU KV tokens. Routed, OpenClaw, and Hermes >300K remain unproven because the legacy routed needle generated a conservative 450,028- token admission estimate and failed 413.

Outcome: direct GPU-only capacity reached 359,900 actual prompt tokens, but routed and client-path requests above 300K remain unproven.

r33 — target-only and batch-token controls

The r33 target-only control adds functional, capacity, and bounded quality evidence on the same released checkpoint:

  • Functional preflight passed 6/6, including 20/20 typed tool calls, streaming tool use, tool-result continuation, and the Responses API.
  • High-reasoning intelligence passed 6/6 attempts, session recall 3/3, and tools 3/3. Low/high/max prompt fingerprints measured 6/85/98 prompt tokens.
  • A single 119,503-prompt-token request measured 17.445-second TTFT, 7,537 effective prefill tok/s, and 73.86 decode tok/s. A short coding smoke passed after each context probe.
  • The engine exposed 15.27 GiB or 283,917 GPU KV tokens. That cannot support a 393,216-token request without another capacity tier. The prepared 393K arm retains FP8 DS-MLA KV and adds 16 GiB native host offload; it was not loaded.
  • Requested context targets were non-monotonic against API-reported prompt tokens. Capacity claims therefore use only the reported actual size, and a harness-integrity ticket remains open.

The matched r33 batch-token arm adds functional, capacity, and bounded performance evidence. Reducing only max_num_batched_tokens from 8,192 to 4,096 lowered peak activation from 1.73 to 1.14 GiB per rank and raised the minimum-rank KV allocation from 15.27 to 15.99 GiB. It passed 6/6 functional checks and the same 119,503-actual-token request. The engine reported 553,243 KV tokens, but the 94.861% reported-token increase does not reconcile with the 4.715% KV-byte increase. A configured 393K start and actual >300K request are therefore still required before claiming over-300K GPU-only capacity.

Outcome: the 4,096-token batch arm reduced measured activation pressure, but the unreconciled KV counters prevent a broader capacity claim.

r16 — baseline DSpark qualification

The r16 revision has functional, capacity, and bounded quality evidence:

  • Low reasoning passed smoke, JSON, three typed tools, streaming tools, tool-result continuation, and Responses. High and max passed smoke, JSON, tools, and tool-result continuation.
  • The context ladder passed 32K, 64K, and a clamped 126,464-token request. The warmed 125,785-prompt-token row measured 19.44-second TTFO, 23.81-second first-visible TTFT, 6,469 effective prefill tok/s, and 128.9 tok/s combined reasoning/visible decode.
  • Nine coding/intelligence/session/tool items ran three times each and passed 27/27 attempts with 4,096 reasoning-headroom tokens.
  • In a same-image c1 A/B, DSpark increased median per-request decode from 64.9 to 130.7 tok/s, aggregate output from 59.6 to 101.7 tok/s, and reduced median E2E from 3.88 to 1.60 seconds.
  • Cumulative DSpark counters recorded 4,865 accepted of 8,830 draft tokens, or 55.1% acceptance and 2.75 accepted tokens per draft.

Outcome: the baseline established repeatable functional, quality, context, and matched DSpark performance evidence at the recorded 128K-era envelope.

r16 — native CPU KV offload

The derived native-offload lane adds functional and capacity evidence:

  • The 128K configuration stored 7.15 GB GPU-to-CPU and replayed 1.13 GB CPU-to-GPU; a 52,495-token replay reached 99.5% cache hits and 0.514-second time to first output.
  • The 256K configuration passed cold 92,754-, 193,064-, and 249,573-prompt-token requests. The largest measured 43.75-second TTFO, 45.58-second first-visible TTFT, 5,705 effective prefill tok/s, and 135.2 tok/s decode.
  • The 8 GiB 256K session stored 11.06 GB GPU-to-CPU, but its follow-ups stayed in the larger GPU prefix tier. With a 16 GiB CPU tier, six distinct 150K planned-context requests stored 13.63 GB; an exact replay then produced 113,408 external hits and loaded 1.002 GB CPU-to-GPU in 0.344 seconds while GPU-prefix hits remained unchanged. Replay TTFO was 0.825 seconds and visible TTFT was 1.974 seconds.
  • The ownership-aware managed lifecycle blocked cleanup while both TP workers mapped the 8 and 16 GiB files, then reclaimed each after container removal and restored split mode.

Outcome: WSL2 native offload and exact CPU-to-GPU replay worked through the measured 256K lane, including ownership-aware managed cleanup.

r16 — GPU-only Pi context

The GPU-only Pi-context lane adds functional and capacity evidence:

  • The 650K/maxseq16 profile recovered a needle at approximately 640K in 120.6 seconds and passed smoke, JSON, three typed tools, streaming tools, tool-result continuation, and Responses.
  • Its matched 32K c1 run completed 3/3 in 16.6 seconds at 5.59-second median E2E, 8,793 effective prefill tok/s, and 141.6 tok/s median decode.

Outcome: the 650K profile passed the bounded Pi and retrieval gates, with the associated latency and concurrency limits preserved below.

2026-08-03 — remote benchmark-worker smoke

The 2026-08-03 remote benchmark-worker smoke adds bounded infrastructure and quality evidence. An 8K native context case passed 1/1, but no larger context bucket was attempted. In the tool-error scenario, the model followed the tool protocol, incorporated results, and retried correctly, then failed the reasoning and final-answer checks. One pinned SWE-bench Verified instance, django__django-11099, resolved under the official grader. This qualifies the remote harness path for a scout campaign; it is not a representative SWE-bench score and does not change the current routing decision.

Outcome: the remote harness path worked for a scout run; the single SWE case is not a representative model score.

1M admission experiments

  • The first 1M profile admitted one sequence and recovered a 985K needle, but a three-tool burst fatally exceeded its locked B12X workspace. It is rejected for Pi agentic use.
  • Raising admission to four retained 1M capacity, passed the same Pi gate and 985K retrieval, and delivered 119.9 tok/s median decode at 32K c1. Near-limit retrieval still took 235.7 seconds.
  • Raising admission again to 16 retained 1,715,610 KV tokens, passed the same Pi gate and 985K retrieval plus a post-probe coding request, and improved median 32K decode to 129.0 tok/s. Later real requests required 703.64 and 687.83 MiB workspaces when only 514.25 MiB was available. The latter had a 19,118-token Pi prompt and only 5,120 requested output tokens, proving that output clamping alone does not make the 1M profile safe.

Outcome: 1M retrieval was feasible, but the retained B12X workspace failures reject these earlier r16 profiles for client-shaped Pi agentic use.

Earlier SGLang and separate vision lanes

The prior SGLang lane used image digest sha256:0aa5324c4f38bc66f4b55e1e12efab821ef614b1a8629259b2810ff72a6570e6, publisher hybrid FP4-expert/FP8 weights, FP8 E4M3 KV, 32,768 context, one admitted request, low reasoning, and no speculative decoding. It passed functional and repeated quality gates. Its final 32K capacity lane completed 11/12 at 2.705-second TTFO, 29.106-second first-visible TTFT, 7,818 effective prefill tok/s, and 11.5 tok/s combined reasoning/visible decode.

A separate vision-adapter package, webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4 (rev 3a8f168c, digest-pinned SGLang v0.5.16, marlin/marlin kernels, TP=2, 4,096 context, --mem-fraction-static 0.97), first-loaded and served on 2026-08-07 with compatibility-only and bounded functional/negative quality evidence: text-lane gates and image conditioning pass, but dense OCR and GUI-affordance reading confabulate against known ground truth, and the checkpoint has no chat template. See the vision first-load finding.

Outcome: retain the SGLang text lane as point-in-time evidence; the vision overlay remains no-promotion because OCR/GUI grounding and chat-template requirements were not met.

External priors and community configuration research

Current external-prior evidence strengthens the research priority without expanding the local contract:

  • DeepSeek reports large agentic gains over Preview, including 82.7 versus 61.8 on Terminal Bench 2.1 and 54.4 versus 7.3 on DeepSWE. The code-agent runs used an unreleased harness at max; two DSBench sets are internal.
  • Artificial Analysis independently scores max-effort 0731 at 50, number 3 of 101 comparable models, while reporting 210 million evaluation output tokens.
  • Community 0731 NVFP4 conversions now exist. MJPansa has the strongest conversion receipt; Auroter has the strongest four-RTX-PRO performance prior. Neither is local TP=2 qualification.

The 2026-08-10 community refresh adds a quality-first decision order without expanding the local contract. It keeps FP8 KV for the first over-300K r33 arm, separates NVFP4 weights from NVFP4 KV, treats Auroter W4A16/FP8-KV as a later precision challenger, and defers W4A4 and NVFP4 KV until matched task-quality gates pass. The proposed 393,216-token target-only arm is an unmeasured experiment design, not a replacement for the recorded r16 profiles. The same refresh also records active vLLM reasoning-template defects and broad Reddit channel coverage; prompt-token fingerprints and a fixed harness are required before any precision comparison.

See the deep research update and its source registry for the benchmark deltas, architecture, runtime matrix, conversion identities, GGUF size ladder, DSpark caveats, and source classifications. See the 2026-08-10 community configuration refresh and its machine-readable candidate ledger for the r33 candidates, precision decision, source classifications, subreddit coverage, and required gates.

Outcome: these sources prioritize future NVFP4 and GGUF experiments; they do not expand the locally qualified contract.

Decision and promotion state

The entries below are dated product decisions. They preserve the human gates and historical promotion state without implying a current live assignment.

2026-08-21–26 — Infernal Invocation r18

Decision: the exact digest-pinned K5 profile was the human-approved llm.primary at 1,048,576 tokens with maxseq8 and batch4,096. The matched no-spec arm is rejected for deployment performance. The guarded router transaction, Mini context/compaction convergence, and real Hermes/Pi/OpenClaw acceptance passed. The profile became historical when Qwen3.8 Flash Next was promoted on 2026-08-26.

2026-08-16 — Infernal Invocation r15

Decision: after explicit human approval, the digest-pinned K5 profile became the exclusive TP=2 text llm.primary at 393,216 tokens. The guarded transaction reran direct context, tools, streaming, tool-result, and Responses gates, installed the router profile, restarted the router, and verified exact post-restart identity and admission. The endpoint-adapted r33 393K profile is the immediate transactional rollback. Martin Vit's upstream receipt qualified 131,072 tokens on native Linux with two RTX PRO 6000 Blackwell GPUs on direct PCIe root ports; this WSL2 393K result is an independent local qualification.

2026-08-02 — r16 maxseq16 promotion record

Decision: the public finding records human approval of the maxseq16 r16 DSpark profile as the exclusive TP=2 Primary for one Pi/OpenClaw coding user, with high reasoning as the client default and llm.rollback preserved explicitly. That is a dated promotion record, not a claim about the live operator assignment. The recorded router's optional per-tier output cap was 32,768 and warned instead of rejecting an oversized caller budget.

2026-08-06 — operator-directed retune

Outcome: the pinned recipe now serves MAX_MODEL_LEN=262144 with MAX_NUM_BATCHED_TOKENS=8192 (previously 650,000/4,096; serve and model names retain the historical 650k suffix). The engine sized GPU KV cache at 272,040 tokens (1.04x concurrency at a full 262,144-token request) and a same-day functional preflight (smoke, structured JSON, tool batch x20, tool-result continuation) passed. The ~640K retrieval, 650K-envelope performance rows, and client smokes above are dated history from the prior envelope and do not transfer without fresh measurement.

2026-08-07 — runtime image upgrade

Outcome: the pinned recipe moved from the r16 to the digest-pinned r27 community image (official 0731 reasoning/tool prompt contract, tiered-offload lifetime fixes, InstantTensor registration fallback, PYTHONHASHSEED=0). KV cache re-sized to 272,107 tokens; a same-day functional preflight passed and a matched 4K/c16 capacity probe showed concurrency parity with the recorded r16 artifact (520 vs 513.5 aggregate tok/s). All r16-labeled performance figures on this page predate the image upgrade.

2026-08-10 — r33 quality control

Decision: the exact released checkpoint passed the target-only/no-spec 131K control on a digest-pinned r33 B12X runtime. The largest validated request contained 119,503 actual prompt tokens. The quality-first 393K candidate keeps FP8 KV and adds host capacity rather than introducing NVFP4 KV before a matched quality A/B. No route or promotion changed.

2026-08-11 — r33 393K promotion

Decision: after human approval, the GPU-only DSpark K5 profile became llm.primary. Direct capacity reached 359,900 actual prompt tokens; OpenClaw and Hermes were updated/restarted at 393,216 context, 32,768 output, and high reasoning. The managed split restoration group is Agents-A1 plus Omni; Qwen was not started. Routed/OpenClaw/Hermes >300K and SWE scoring remain open.

2026-08-10 — batch-token A/B

Outcome: the otherwise matched 4,096-token arm reduced profiled activation pressure, increased GPU KV allocation, passed the same bounded gates, and was healthy/direct-only at campaign close. It is the preferred basis for a GPU-only 393K experiment, but no route or promotion changed and no request above 300K has yet run.

Remaining gates and historical scope

Earlier r16 1M profiles are experimental only; that statement does not apply to the exact qualified and promoted r18 profile. The current exclusive AI-only policy has no separate video-workload reserve gate; the historical 3 GiB reserve failures below remain point-in-time evidence from mixed-use policy. Remaining gates include sustained multi-turn high/max testing, fixing and requalifying the client-shaped 1M B12X workspace failure, and pinned 0731 NVFP4 W4A16/W4A4 comparisons.

2026-08-07 — vision-adapter first load

Decision: no-promotion, evaluated separately from the text Primary. The WebBrain 0731 vision overlay loaded and served on TP=2 with grounded image conditioning, but OCR/GUI reading confabulated and the checkpoint exposes no chat template, so it cannot serve a router chat client. The 650K Primary was restored and its health verified in the same session; no alias, route, or promoted serve changed.

Failures and gotchas

Startup translation and KV admission

The first r16 start exposed an offline nested-speculative-model localization gap. The second exposed native-Linux NCCL defaults incompatible with WSL2. The third loaded both ranks but missed the 131K KV admission gate by 0.15 GiB at max_num_seqs=16; the qualified c1 recipe reduces admission to eight and the CUDA-graph cap to 48 without raising its memory-utilization ceiling.

Historical reserve and workspace pressure

Both r16 profiles fail the 3 GiB reported-free reserve. Per-context sampling found only 1,179-1,203 MiB free on dark-compute-a and 2,031 MiB on dark-compute-b. WSL/WDDM global allocation differs from native Linux, but a successful request does not authorize silently weakening the gate. That gate was tied to mixed graphics/video operation and is not part of the 2026-08-10 exclusive AI-only r33 acceptance policy.

Moving Windows display output to the AMD iGPU allowed the previously failing maxseq16 graph envelope to start. The 650K profile reported only 797/805 MiB free after its workload, the 1M/maxseq4 profile reported 207/209 MiB after the 985K probe, and the 1M/maxseq16 profile reported 339/335 MiB. The experiment intentionally waived the separate 3 GiB acceptance gate while retaining GPU_MEMORY_UTILIZATION=0.975; this is a capacity result, not a new reserve policy. The 1M/maxseq1 profile also remains an explicit retained failure because its locked 514.25 MB B12X workspace could not satisfy a later 873.62 MB compressed-MLA allocation.

Reasoning-budget exhaustion

Reasoning-budget exhaustion is still a real operational risk. One earlier SGLang capacity request and one additional r16 no-spec control run completed without visible content. High/max support proves protocol compatibility, not that arbitrarily large reasoning budgets are efficient.

Redistribution boundary

The pinned runtime source repository has no root license file at the tested revision. The model weights are MIT, but local qualification does not establish permission to redistribute the image or derived runtime code.

Historical-invalid attempts

The historical NGC vLLM architecture rejection and aborted NVFP4 load are not measurements of the current 0731 checkpoint.

Automatic rollback gap

The r18 promotion also reproduced the existing automatic recovery defect for exclusive-to-exclusive rollback: ordinary bring-up correctly refuses the rollback serve without an ownership transfer. The r33 recovery profile remains available through the managed mode transition; automatic promotion rollback is not live-proven until the P1 orchestration issue is fixed.

Native-offload shared-memory lifecycle

Native offload originally crashed on WSL2 because CUDA host registration of the process-shared mmap conflicted with the allocator lifetime. The derived image skips registration only for that mmap while preserving the global V2 UVA path. A separate 256K failure was stale tmpfs exhaustion: four orphan offload files filled /dev/shm, producing mmap.madvise: Bad address. Anvil Serving now performs two ownership checks and exact-path cleanup before load and after managed teardown. Page-cache reclaim alone is not sufficient.

Dated run history