Skip to content

DeepSeek V4 Flash 0731 Infernal Invocation r15 393K promotion

Date: 2026-08-16

Decision: human-approved llm.primary; r33 393K is the managed rollback

Evidence: functional, capacity, matched performance, bounded quality, routed acceptance

Outcome

The digest-pinned Infernal Invocation r15 profile is qualified on the local dual-RTX-PRO WSL2 host at a 393,216-token configured window and is installed through the managed serving path. It owns both equal cards in exclusive TP=2, the router alias llm.primary resolves to its exact served identity, and the route is healthy and admitting. The prior r33 393K profile remains the transactional rollback on the same fixed endpoint port.

This promotion does not transfer the upstream author's qualification. The upstream receipt covered 131,072 tokens on native Linux with two RTX PRO 6000 Blackwell GPUs attached to direct PCIe root ports. The 393,216-token WSL2 result below is an independent local qualification.

Source and author credit

This work was inspired by and translated from Martin Vit's (voipmonitor) Infernal Invocation r15 work in local-inference-lab/rtx6kpro at 55323f94 and local-inference-lab/blackwell-llm-docker at 2c301121. The image composition also incorporates the contributors to local-inference-lab/b12x PR #214 and local-inference-lab/vllm PR #320. The public recipe references these sources and does not redistribute their runtime source or image. Neither pinned repository exposes a root license file, so local use does not establish redistribution permission.

Immutable configuration

  • Checkpoint: deepseek-ai/DeepSeek-V4-Flash-0731 revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb.
  • Image: voipmonitor/vllm@sha256:f1b13c8604b274212e1164def7d4ed7a4cac9e4f7fa06fa1739730195eca4e18.
  • Runtime: vLLM integration tree 068fc8e7270b92077ba753d002da179c865e444d, B12X tree 96e5d3d5c2057fa5d4f542e2368951ddbdcb5b42.
  • Hardware: two RTX PRO 6000 Blackwell Max-Q Workstation Edition cards, Windows 11 Docker Desktop/WSL2, PCIe without NVLink, exclusive TP=2/DCP=1.
  • Serving: B12X W4A8, FP8 compressed MLA KV, InstantTensor BUFFERED, 393,216 tokens, maxseq8, batch4096, fixed probabilistic DSpark K5.
  • Offload: native KV offload, native L2, and LMCache disabled.
  • K5 recipe and matched no-spec control.

Matched speculative-decoding A/B

The control retained the exact image, checkpoint, TP/DCP layout, context, batching, maxseq, FP8 KV, allocator, loading, memory utilization, and WSL2 controls. Only DSpark, the loopback port, and served evidence name differed.

Shape No-spec median decode K5 median decode Change No-spec / K5 effective prefill
4K, c1, 3 runs 76.4 tok/s 150.0 tok/s +96% 8,237 / 8,741 tok/s
32K, c1, 3 runs 76.767 tok/s 119.245 tok/s +55% 9,262 / 9,071 tok/s

Both arms passed the full protocol gate. K5 speculative counters sampled 72.8% acceptance in the principal lane; later long requests varied, so the speed result remains prompt- and trajectory-specific.

Correctness, context, concurrency, and quality

  • Direct K5 passed smoke, JSON, retrieval, tools 20/20, streaming tools, tool-result continuation, and Responses.
  • Direct retrieval passed at 118,080 and 351,118 actual prompt tokens; the largest completed in 53.133 seconds and returned the exact needle.
  • The authenticated router passed a calibrated 340,119-actual-token request in 58.891 seconds and recovered its needle through llm.primary.
  • Short concurrency completed 8/8 at c8 with 267 aggregate output tok/s.
  • Long concurrency completed 2/2 at c2 with 99,175 prompt tokens per request after applying the qualified 4,096-token reasoning headroom.
  • Repeated tools, session recall, unified diff, and parallel timeout triage passed 12/12 attempts.
  • The engine reported 797,689 KV tokens, or 2.03 configured full windows. The router therefore admits at most two requests; this is a capacity ceiling, not a broad 393K-c2 quality claim.

The sanitized machine-readable summary is summary.json.

Serving-path acceptance and rollback

The guarded transaction quiesced and drained primary-local, reused the already-qualified exact-identity target, reran the complete direct 390K/tools gate, atomically installed the router profile, restarted the router, and verified post-restart identity readiness. The route then passed coding, JSON, tools 20/20, streaming tools, tool-result continuation, Responses, and the calibrated 340,119-token request.

OpenClaw-compatible Anthropic Messages calls through llm.primary returned an exact readiness marker and a valid typed workspace_status tool call. This is wire-path evidence, not a fresh actual Mini OpenClaw client run: the reachable Mini controller does not implement the current OpenClaw status tool. No Mini configuration was changed.

The rollback profile is the most recent working r33 B12X/DSpark K5 393K configuration. Its checkpoint, image, context, batching, maxseq, speculation, and offload policy remain unchanged; an endpoint-only adapter moves it from 39075 to the transaction's fixed 39077 port. Recipe dry-run, manifest lint, rollback-check, and promotion dry-run all passed before the live cutover.

Retained failures and caveats

  • The generic thinking-disabled request did not disable r15 reasoning. Default thinking is therefore the qualified contract; the router does not advertise caller override.
  • The stock nominal-390K English preflight was rejected by the router with HTTP 413. Its conservative byte/4 estimate exceeded 393,216 even though the direct server measured 351,118 actual tokens. A token-dense calibrated routed probe passed at 340,119 actual tokens; the 413 is retained rather than relabelled as success.
  • The first 128K/c2 attempt allowed only 512 completion tokens, so both requests finished reasoning without visible content. Repeating with the qualified 4,096-token budget passed 2/2.
  • Startup warned that K5 reduced max_num_scheduled_tokens to 4,064. No exception, OOM, or request failure followed; no unmeasured tuning was mixed into the A/B.
  • Idle reported free VRAM is about 1.5 GiB per card. This exclusive AI profile does not retain a graphics/co-resident reserve.
  • The managed restart and rollback commands now reference the canonical public checkout at the merged recipe revision. This closes the earlier linked- worktree cleanup risk; the private operator-state merge itself did not restart the already-running model or router.