Skip to content

DeepSeek V4 Flash 0731 Infernal Invocation r18 1M qualification and promotion

Date: 2026-08-21 Decision at publication: human-approved live llm.primary; the guarded promotion, authenticated routed gate, Mini generation-2 convergence, and real Hermes/Pi/OpenClaw acceptance passed. Measured hardware: two NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition cards in exclusive TP=2 over PCIe without NVLink, under Windows 11, Docker Desktop, and WSL2.

Result

Martin Vit's Infernal Invocation r18 recipe was translated to the managed Anvil Serving lifecycle, pinned byte-for-byte at its upstream revisions, and qualified locally with the requested 1,048,576-token context and fixed probabilistic DSpark K5 configuration. The target passed the complete direct functional gate, repeated deterministic agentic checks, a 160-call structured tool soak, client-shaped reserve probes, short and long concurrency, calibrated retrieval through 1,040,063 actual prompt tokens, a matched no-spec control, and a clean post-reload gate.

The engine reported 1,323,176 GPU KV tokens, or 1.26 full configured windows, with eight admitted sequences. Router admission is intentionally bounded to one full-window request at a time. Native KV offload and LMCache are disabled.

After explicit operator approval, the fixed-port K5 variant was promoted through the managed exclusive-mode transaction. The live owner is healthy at TP=2, and the authenticated router reports the exact expected and observed r18 identity ready and admitting for llm.primary. Auxiliary, secondary, voice, OCR, general-vision, and video aliases were preserved rather than folded into the new text route.

Immutable upstream identity and credit

  • Source recipe: ds4dspark-infernal-invocation-r18.md, local-inference-lab/rtx6kpro commit 12272507e64752af7a6cde28540edf8ee6c20e59.
  • Compose source: docker-compose-ds4-infernal-invocation-cu133-r18.yml, local-inference-lab/blackwell-llm-docker commit 203e1a11a373a0486897a5a284409ba4417d0f7a.
  • Upstream author/publisher: Martin Vit (voipmonitor). The local result is an independent WSL2 qualification and does not transfer or expand the upstream native-Linux receipt.
  • Model: deepseek-ai/DeepSeek-V4-Flash-0731 revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb.
  • Runtime image: voipmonitor/vllm:infernal-invocation-vllmf0fa1ce-b12x75787c7-fi1ac6942-cu133-torch213-20260818-r18, registry digest sha256:414ec7d0d28358cfd8af0697f330f5c8acbb80e4dc4e5ba69c9fd5b5855ea804.
  • Runtime trees: vLLM f0fa1cefc1865d316c2478525f550e7646addc40, B12X 75787c7a7431b3bea414d2ebf5f2b8671b23eb33, and LMCache e045d729bc5c4c63a40e13d032f42923de97812f.

The upstream receipt physically qualified a 262,144-token configuration. Its 1M scheduler shape was therefore an external recipe prior, not proof of local 1M execution. This finding supplies the missing local full-window and client-shaped evidence.

Exact translated configuration

Layer Qualified value
Target B12X W4A8 with FP8 compressed MLA KV and FP32 sliding-compressor state
Parallelism TP=2, DCP=1
Context MAX_MODEL_LEN=1048576
Scheduler MAX_NUM_SEQS=8, MAX_NUM_BATCHED_TOKENS=4096
Memory GPU_MEMORY_UTILIZATION=0.975; model load 81.09 GiB per rank
Speculation fixed probabilistic DSpark K5; FULL target, draft, and context-KV graphs
Transfer InstantTensor BUFFERED
Cache prefix cache enabled; native KV/L2/LMCache off
WSL2 direct NCCL P2P and NCCL cuMem device/host allocation disabled; expandable segments disabled

The checked-in K5 recipe and matched no-spec control pin the complete managed configuration. The fixed-port promotion variant changes only the loopback endpoint from the isolated qualification port to the existing r33 rollback port; model, runtime, context, scheduler, KV, graph, and speculative settings are unchanged.

Functional, quality, and tool gates

  • Direct preflight passed coding smoke, structured JSON, a 32K needle, tools 20/20, streaming tools, tool-result continuation, and the Responses subset.
  • Repeated deterministic quality passed intelligence 6/6, session 3/3, and tools 3/3 with zero recorded failures at 512 visible tokens plus 4,096 reasoning-headroom tokens.
  • Eight additional shared-prefix batches passed 160/160 structured tool calls. A bounded post-soak log window contained zero xgrammar/FSM, traceback, OOM, HTTP 4xx/5xx, or error-level entries.
  • After the matched no-spec control, K5 was reloaded through the managed exclusive-mode transaction and again passed smoke, JSON, tools 20/20, streaming tools, tool-result continuation, and Responses.

Context and client-shaped capacity

Workload Actual prompt tokens Result Request time
calibrated retrieval 354,010 pass 60.882 s
calibrated retrieval 585,118 pass 92.087 s
calibrated retrieval 900,118 pass 175.999 s
near-limit calibrated retrieval 1,040,063 pass 187.347 s
Hermes-shaped reserve, 20K / max 5,120 14,357 pass retained privately
OpenClaw-shaped reserve, 32K / max 8,192 23,453 pass retained privately

The near-limit request returned the exact needle with a visible answer and finish_reason=stop, leaving approximately 8.5K tokens of combined context and output headroom. It proves local execution near the declared window; it is not a broad 1M reasoning-quality score.

Performance and matched speculative A/B

All A/B rows use the same checkpoint, image, TP/DCP, context, scheduler, batching, KV format, and request shape. Only DSpark K5 is changed.

Workload K5 median decode No-spec median decode K5 median E2E No-spec median E2E
4K, c1, 3 requests 142.1 tok/s 76.4 tok/s 1.576 s 2.350 s
32K, c1, 3 requests 129.5 tok/s 76.3 tok/s 4.098 s 4.660 s

K5 improved median decode by 86.0% at 4K and 69.7% at 32K. Median effective prefill was 8,579 versus 8,438 tok/s at 4K and 8,993 versus 9,287 tok/s at 32K; the 32K result therefore attributes the end-to-end win to speculative decode rather than a prefill claim.

Additional K5 capacity completed 8/8 at c8/4K with 272.33 aggregate output tok/s. A long c2 run completed 2/2 at 490,861 prompt tokens per request with no preemption or workspace failure. These are bounded concurrency results, not a claim that eight simultaneous full-window requests fit.

Live promotion, clients, and rollback boundary

The operator explicitly authorized validation and promotion. Qualification selected the K5 arm and rejected the no-spec arm for deployment performance. The live acceptance then established:

  1. the exact r18 serve owns both cards in managed exclusive TP=2 mode and reports healthy;
  2. authenticated llm.primary routing has exact expected/observed identity and passes coding smoke, structured JSON, a 32K needle, tools 20/20, streaming tools, tool-result continuation, and the Responses subset;
  3. Mini generation 2 declares the Primary context as 1,048,576. It keeps auxiliary, secondary, and vision context at 131,072, and voice context at 32,768 without changing protected provider or authentication fields;
  4. Hermes retains 50%/20% compression and its 5,120-token output ceiling; Pi retains a 20,000-token recent window and 16,384-token reserve; OpenClaw retains safeguard compaction, 50,000-token reserve/floor, 30,000 recent tokens, and a 50% history-share ceiling;
  5. the affected gateways restarted cleanly, Hermes completed an exact terminal sentinel with exit zero and no raw DSML, Pi completed through its normal extension-loaded process, and OpenClaw completed through the running gateway with no fallback.

The exact r33 393K fixed-port profile remains the declared recovery target and can be restored through the managed exclusive-mode transition. The automatic exclusive-to-exclusive rollback inside serves promote is not claimed as proven: a failed gate reproduced the existing P1 recovery defect, where ordinary bring-up is correctly rejected because ownership was not transferred. Exclusive admission remains intact; orchestration must be fixed.

Public evidence

The sanitized summary contains the immutable identities and decision-relevant aggregates. Raw request bodies, prompts, private topology, GPU identities, route URLs, and logs remain private operator evidence.

Caveats

  • The result is exact to this checkpoint, digest, runtime trees, WSL2 translation, two-card topology, DSpark K5 policy, and scheduler shape.
  • Reported post-load free VRAM is under 0.5 GiB per card. The profile is an exclusive AI workload and does not establish a co-resident graphics reserve.
  • A startup-only PyTorch compile diagnostic logs deprecated enum constant registration at error level. The post-soak request window is clean; this is retained as a runtime caveat rather than misreported as a request failure.
  • The 160-call tool soak targets the prior xgrammar/FSM regression but cannot prove absence of every future structured-output edge case.
  • Full-window concurrency is one. Short c8 and long c2 evidence do not raise that router admission ceiling.
  • The declared r33 rollback is a managed recovery path, but the promotion command's automatic exclusive-to-exclusive rollback remains an open P1 and must not be described as live-proven.