[JJ r1] DeepSeek-V4-Flash source merge contract #94

Purpose

This issue is the source merge contract for the qualified
DeepSeek-V4-Flash-0731 Jovian Judgement r1 image:

voipmonitor/vllm:jovian-judgement-vllm28bc825-b12x8a5b9bf-fi1ac6942-cu133-torch213-20260903-r1
registry digest: sha256:9e8cbbe80423510ad77b35fc7414101e2a5d4529ccf6d8c7ee08ab874483bac7

The image is built from pinned public base revisions plus the pull-request
heads listed below. Merging those heads makes the qualified runtime
reproducible from canonical branches without changing its source tree.

Runbook:
https://github.com/local-inference-lab/rtx6kpro/blob/master/models/ds4dspark-jovian-judgement-r1.md

vLLM: dev/jovian-judgement

• Register speculative CUDA graph row counts during B12X warmup vllm#628 at
  cbb66bdff1763c174ebc794a7f968930e956580f

  Registers scheduler-reachable speculative row counts before B12X CUDA-graph
  warmup. DSpark graph capture can therefore compile the row contracts that
  production replay will execute instead of discovering an uncovered contract
  after startup.

• fix(ds4): clear custom all-reduce selectors in NCCL mode vllm#630 at
  5b6fb80f5c868b62da2c01c3f52861b34c84d8ac

  Makes ALLREDUCE_MODE=nccl remove B12X backend and cutoff selectors from the
  child environment. Explicit NCCL selection therefore cannot inherit a
  custom all-reduce policy from the parent process.

Integration tree:
28bc825a1321bb480fc3294179fd34afeb468389

B12X: master

• fix(comm): make TP2 graph peer-push generation-safe b12x#246 at
  2a2baaf3979c1f47f8fd139aab5c6eeea35003e4

  Replaces the TP2 CUDA-graph peer-push sentinel with generation-tagged
  producer/consumer publication. Replayed graph generations cannot consume a
  readiness value left by an earlier replay.

• Use a valid W4A8 MoE profiling oracle b12x#302 at
  f8a7b9c4070754ed4399c71fc4306d8970711e1c

  Profiles W4A8 routed-expert candidates against an oracle that supplies valid
  MXFP8 scale and route metadata. Candidate qualification therefore measures
  the declared operation instead of an invalid synthetic call.

Integration tree:
8a5b9bfbf59ad61d87efdc8017b91a269d5a319c

LMCache: release/v0.5.2-glm52-dcp-base

• fix(vllm): support engine-driven hybrid KV transfers LMCache#44 at
  97ede799d6605ca1bd5285582df4e74a3d3c7b0d

  Transfers interleaved 64-head blocks-first KV pages with a legal CUDA launch
  geometry and the tensor's physical dimension-zero stride. This is required
  for DeepSeek V4 compressed-MLA cache groups in the engine-driven LMCache
  path.

Integration tree:
eb4c227f68a4e1c45d6b8edf6b4934e18f6d1f8b

The LMCache code-quality workflow reports pre-existing mypy failures in files
outside PR #44's diff. The pull request's focused CUDA suite reports 114 tests
passing, and the exact image passed the live store/evict/replay gate described
below.

Qualification Evidence

Hardware: two RTX PRO 6000 Blackwell GPUs connected through one PCIe switch.
Runtime: CUDA 13.3, PyTorch 2.13.0, NCCL 2.31.2, TP2/DCP1, B12X W4A8, FP8
compressed MLA KV, fixed probabilistic DSpark K5.

Scope Boundary

Native vLLM filesystem L2 is not part of this release contract. The Compose
profiles default to GPU KV storage and enable host-side KV movement only when
LMCACHE_MODE=ram or LMCACHE_MODE=disk is selected. vLLM PR #631 is an
independent optional implementation and is not required to reproduce Jovian
Judgement r1.