Add DeepSeek-V4-Flash-Vision-Exp recipe and runtime dependency mod

[isolitude]

Summary

Add a cluster-only recipe for DeepSeek-V4-Flash-Vision-Exp (FP8) on a dual DGX Spark cluster, with unified vision indexing + DSpark speculative decoding, plus the runtime dependency stack it needs.

What's included

• recipes/deepseek-v4-flash-vision-exp.yaml — cluster-only recipe. Builds the vllm-node-b12x image with --exp-b12x --rebuild-vllm --apply-vllm-pr 634 (the vision feature is not merged into dev/jovian-judgement), applies the b12x runtime overlay + LMCache dev+5 install at container start.
• mods/jj-ds4-vision-deps — the container-start runtime stack:
  1. b12x source overlay (pure Python/JIT, no rebuild): PR #301 FP8 dual-cache prefill + sparse topk-512, PR #246 TP2 comm/pcie, PR #306 mHC rms_eps=1e-20 allowance.
  2. LMCache dev+5 build/install from a vendored tarball (lmcache-dev5-src.tar.gz, ~2.2 MB): dev base @7ed46754 + follow-up PRs #49/#50/#51/#55/#56 — this replaces the previously-depended-upon PR #44, which vLLM PR #634 dropped.
  3. flashinfer — no-op (baked 0.6.18 already carries the SM120 topk-512 dispatch).
     run.sh is idempotent and self-verifying; verify.sh checks the three claims against the installed environment.
• Docs — docs/deepseek-v4-flash-vision-exp.md (dependency stack, PR #634 status, caveats) and a CHANGELOG entry.

Dependency stack

Component	Patch	Where
b12x	PR #246 TP2 peer-push + PIECEWISE binding	mod (runtime overlay)
b12x	PR #301 FP8 V4 dual-cache prefill, sparse topk 512	mod (runtime overlay)
b12x	PR #306 rms_norm_eps=1e-20	mod (runtime overlay)
vLLM	PR #634 Vision (NOT in branch)	baked via --apply-vllm-pr 634
vLLM	PR #553 engine-driven LMCache + expandable CUDA	already in dev/jovian-judgement
vLLM	PR #671 fused padded-query KV accounting	already in dev/jovian-judgement
LMCache	dev + PRs #49/#50/#51/#55/#56 (replaces PR #44)	mod (vendored install)
FlashInfer	803c4664 SM120 sparse-MLA topk-512	baked in 0.6.18 (no-op)

Why LMCache dev + follow-up PRs, not PR #44?

PR #634's runtime-dependency list dropped PR #44 (the release/v0.5.2-glm52-dcp-base lineage) in favor of five dev-lineage PRs #49/#50/#51/#55/#56. The dev engine-driven path is a different implementation from #44's (no EngineDrivenContextPickle; the context lives in worker_transfer.py as EngineDrivenTransferContext). Build the dev+5 source, not #44 or any release/v0.5.x wheel.

Notes

• --hf-overrides '{"architectures":["DeepseekV4ForConditionalGeneration"]}' is required (the model loader resolves from raw hf_config.architectures).
• Weights must be present in the HF cache of both nodes.
• This is experimental (as are b12x builds in general); keep the repo updated.