ProprietaryLegal/dsv41-sampled-drafting

DSV41 sampled drafting (PP4): sampled DSpark drafts with exact target-side acceptance

DeepSeek-V4.1-Flash's DSpark drafter normally proposes greedy drafts. Its acceptance law, sum of min(p_target, p_draft), favours drafts sampled from the drafter's own distribution. PP4 moves the remote DSpark drafter (on two NVLinked RTX 3090s) to sampled drafts and carries the draft probabilities to the target's verifier, so the output distribution stays exactly the target's. Snapshot of the private offload/pp4-win2 lane at 81969f70009a (2026-10-27); full tree lane/pp4-win2 (and the promotion package lane/pp4-promote) in dsv41-offload.

Full source: dsv41-offload · Drafter worker: dsv41-remote-dspark-drafter-3090 · Hub: pli-labs

Published by PLI Labs, the open research arm of Proprietary Legal Intelligence. PLI builds AI systems around a legal organization's own work product, combining retrieval, staged workflows, validation, and attorney review. Visit proprietarylegal.ai to explore our work and discuss what your organization needs.

Numbers

	Value	Label
Accepted drafts per round at fixed depth 5, greedy → sampled (offline replay)	2.092 → 2.306 (+10.2 %)	MEASURED (offline)
Committed tokens per round, (1 + accepted)	3.092 → 3.306 (+6.9 %; 95 % interval on the difference +5.7 to +8.2 %)	MEASURED (offline)
Sampling selector cost	+0.31 ms per round at 1 user, +0.79 ms at 4 users	MEASURED
Net single-user decode gain after the selector cost	+3 to +6 %	ESTIMATED
End-to-end decode on hardware	not yet measured against the adaptive production schedule	—

Correction: the 2026-10-27 version of this card gave +10.5 % accepted and +7.1 % delivered; those rounded 2.092 / 2.306 before dividing. The exact figures are above.

What is in code/

• offload/dspark/probabilistic/: the sampled-draft path, wire format and offline replay.
• offload/dspark/worker/sampled*.py, sample_policy.py: worker-side sampling (fused).
• offload/dspark/spark/remote_dspark/sampled_wire.py: Spark-side transport of draft probabilities.
• offload/audit/verifier_rng/: a gate for the verifier's random draws on PP4's sampled path (vLLM #61161 class of bug).
• offload/launch/windows/pp4_*.py and tests: admission (physical pair qualification, greedy near-tie gate with a registered 1.1-nat ceiling, KL probes) and the v2 measurement, including a production-policy mode that compares PP4 against production's own adaptive depth schedule. The primary metric is tokens per server decode second, because PP4 deliberately spends a little time per round to accept more drafts.

Status and tests

Admission on hardware: round 16's fix to the admission proxy was confirmed on hardware on 2026-10-27; the fix for the admission API semantics (81969f7, included here) is under independent review before the next admission window. No live speed measurement yet. Not promoted to production.

Tests: 529 passed, 146 failed or errored (distinct tests), 28 skipped (CPU only, no network). Of the failures: 84 also fail on the private source commit in the same sandbox (pre-existing: GPU, serving hosts, deployment state, git history a single-commit tree lacks, or an existing failure); 0 pass when their file runs alone; 54 need data on our hosts; 8 need files of the full tree (they pass in dsv41-offload). Per-branch results for the full tree are in reports/tests/TESTS.md of dsv41-offload (lane/pp4-win2).

Host names are role names (sparkhead = rank 0, sparkworker = rank 1, x86host = the x86 offload host, controller = the workstation running the harness); our directories appear as neutral absolute paths under /opt/dsv41.

Licence

Apache-2.0 (LICENSE); third-party notices in NOTICE and LICENSES/.