[yifjiang]

Enable opt-in reuse of a longer completed prefix retained by a decode worker during a later prefill. Bounded donor lookup is separate from execution selection; unavailable donors, cancellation and incomplete delivery preserve fallback. The feature is disabled by default and does not modify production vLLM.

Each role’s manager owns one TP1 bank or rank-local TP2 banks sharing logical slot IDs. Atomic descriptors bind geometry, GPU identity and owner lifetime. Inventory becomes ready after every worker acknowledges its store. Workers import their same-GPU bank, with source pins and destination storage protected through native quiescence.

CUDA IPC is default. Explicit MNNVL uses CUDA FABRIC and isolated UCX GPU RMA, TCP control and no CPU payload fallback. Scope: raw TP1/TP2, local multiprocessing, DP=PP=CP=1. The new D-to-P RDMA profile and native peer-death recovery remain unsupported. Companion typed hints/multipart delivery and experimental native patches: KVCR #104. Part of #16747 and #16746.

Source and validation

Public head be2fbeba includes main cb5f9a67, with exercised source inventories unchanged. Actual ARM runtime/components producer remains 0c8cea0c, KVCR 83e5e9e6, image sha256:f727f1274998d6fdd40c8147cd156e2e190157c36be220c46cff0ae869bb3451. Source equivalence does not relabel the benchmark producer.

• Allocated producer gates: router 1,427, LLM 3,828, both admission integrations; Python feature 190, runtime arguments 44, TCP admission 58. Format, selected Clippy, release/package/native/import QA pass. Separate public be2 CI: 84 successful /43 skipped, no failed or pending reported checks.
• Fresh GB200 native oracle: positive 7 keys /229,376 B /14 PtoP copies; missing-key 6 keys /196,608 B /12 copies, untouched missing rows, no false completion. Both ranks, guards, pins, close and zero host payload pass. Bank tests: 19 mocks plus one real two-GPU case.
• Earlier coherent 1.7: four real GPU adapter/IPC cases; Qwen's seven requests/twelve traces pass under batch-invariant FLASH_ATTN/FA2, with complete OFF/ON/fallback IDs. ON delivers 37,748,736 B, computing 18 versus 274 tokens. The default-backend mismatch is preserved.
• An independent MiniMax 4,096→1,024 pilot passes complete raw/counter replay.

Fresh MiniMax on SLURM GB200 TP2	Actual result
OFF4 and profiled reuse correctness	Full raw IDs/counters pass; six-actor second-turn proof delivers 136,249,344 B, equally across ranks, with zero host payload
Original three-sample normal-policy calibration	NEGATIVE: headroom −59.479 ms, no positive token coefficient or calibrated mode
New adaptive forced diagnostic: six balanced pairs	TTFT 2.90% slower in 6/6 pairs (median +12.00 ms); completion time mixed and essentially flat; fixed control coefficients, not calibrated gain

The adaptive comparison was separately declared after negative calibration and before its own first paired request; it was not predeclared before overall traffic. Prior STOP failure and the separate DGD counter/expired-authentication block remain preserved. SLURM evidence is not a DGD result. No MiniMax calibrated gain is claimed. This PR remains draft.

No performance gain was observed on this tested MiniMax workload. All reported latencies are second-turn client observations; profile/calibration timings are excluded. After all 36 requests completed, prefill exited 1 during post-finish SIGTERM/cancellation; both owned model groups stopped and fresh process-absence checks passed.

Exact sources, commands, all six pairs and preserved failures are in the reproduction record.

[devin-ai-integration]

✅ Dynamo PR CI passed — run 41178467317 (attempt 1) on be2fbeba95

Gate checks: ✅ backend-status-check · ✅ deploy-status-check · ✅ dynamo-status-check

_{Posted automatically by Devin for run 41178467317. Updated on every full-CI run of this PR.}

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a190c1-2384-728b-bb21-ce12727d0694.docs.buildwithfern.com/dynamo/dev