| # PP4 integration (default OFF) |
|
|
| `DSPARK_SAMPLED_TOP32=1` selects the new K2 worker path; the experimental |
| engine still requires `VLLM_DSPARK_SAMPLED_TOP32=1` and PP3's guarded standard |
| verifier ABI. No production launcher enables either switch. |
|
|
| `TPDrafter.heads` branches before the original selector. A sampled position |
| launches one `sampled_fused._kernel` per physical rank, or one kernel containing |
| both logical vocabulary shards in solo emulation. It reads the existing BF16 |
| head logits and the matching Markov cache/weights directly. Cached rows add |
| their bias; uncached rows retain the per-column FP32 product/tree reduction |
| and BF16 bias/add rounding. No additional conditioned full-vocabulary tensor, |
| greedy selection, full-logit peer copy or full-logit host copy occurs. |
|
|
| Each tile writes only its top32 packed value/ID candidates. A GPU release/acquire |
| retirement chain selects the last retiring CTA to finish the row; it never |
| spins for unscheduled blocks. That CTA reduces candidates, normalizes top32 |
| at tau0.7, allocates exactly2^24 tickets and draws with the existing request-owned |
| Philox key/round/absolute-position domain. IDs, emitted FP32 q and sampled |
| tokens retain `dyadic24-floor-first-v1` semantics. |
|
|
| On the physical pair, each final CTA sends32 packed candidates (256B/request/ |
| position), signals with a system release and waits with a bounded system |
| acquire. Both ranks reduce the same union, draw from the same q/key and |
| acknowledge consumption before a subsequent position can reuse the inbox. |
| Epoch exhaustion, timeout and bad logits fail closed with sticky peer errors. |
| The host discards the entire reply. An invalid intermediate uses an in-range |
| internal continuation0 while retaining its error flag, so later Markov |
| positions cannot dereference-1. This internal value is never a successful |
| proposal or a substitute for q. A sticky peer failure requires worker teardown. |
|
|
| Peer addresses are runtime arguments, not JIT constants. Both rank |
| specializations are compiled before either kernel launches; otherwise a GPU |
| could timeout while its peer was still compiling. Capture refuses an unprimed |
| paired specialization. Physical-pair execution/timeout visibility has not been |
| qualified by solo compilation and remains a coordinator gate. |
|
|
| The worker captures separate sampled graphs alongside original graphs when |
| the switch is explicitly enabled. Graph inputs include per-rank request keys |
| and T>0 masks; outputs/q/error state have fixed lifetime per(B,k,n_out). |
| Missing sampled graphs on a graphed worker refuse eager fallback. Fixed |
| depth4/5 is supported, including production query-block5/n_out4 and5; unused |
| output/confidence/probability padding is zero. The normal server uses its |
| negotiated fixed block depth. |
|
|
| OFF uses the original selector and graphs. An all-T0 request batch with the |
| switch ON reuses the original graph and output objects. Mixed batches use |
| the fused selector with a per-row T0 argmax and no Philox draw; q is zero on |
| those rows. Mixed T0 equality therefore needs numerical validation in addition |
| to the all-T0 branch test. Native, partial-rank, margin and traced sampled |
| head execution are outside this path and refuse. |
|
|
| PP3's transport, generation/slot mapping, probability validation, standard |
| min(1,p/q) acceptance, full-target residual and failed-proposal invalidation |
| remain unchanged. Frozen heldout replay consumes the actual new producer's q. |
| Solo qualification and whole-worker timing do not establish physical-pair, |
| GB10, grammar/parser or live throughput qualification. The separate PP4 |
| coordinator runner refuses missing hash-bound admission evidence. |
|
|
|
|
| ## Round2 review corrections (CPU only) |
|
|
| S52's four fixes are implemented in the PP4 launcher/preparation layer. The |
| nested production-imported remote_dspark package is replaced exactly once. |
| Verifier preparation starts from the reviewed full K1 runner682eb895…; |
| verifier-k1-receipt.json records refreshed source/output hashes. Both MTP None |
| guards remain. Receiving-host ownership includes nested experimental/recovery |
| SSH/SCP, with persistent fences, leases and deferred exact-HOLD close. Failed |
| rounds gate growth over an immutable post-warmup/post-greedy baseline across |
| headline, profiled and final snapshots. |
| |
| Physical qualification has a concrete coordinator-only pair_qualify entrypoint, |
| called first inside each HOLD under receiving-host registration and both GPU |
| locks. It cannot synthesize qualification from these solo receipts. It executes |
| both peer kernels, changed-input graphs, mixed T0, isolated timeout/sticky |
| teardown, and physical whole-worker identity against the archived solo N4 |
| receipt. It was written and CPU rehearsed, **not executed on hardware**. |
|
|
| The whole_worker script now exposes a guarded physical qualification-only |
| mode. full-worker.json remains the historical56f2515 solo measurement; its |
| runner hash refers to that historical source, not the new wrapper. Worker |
| numerical code and historical measured receipts were not changed or rerun. |
| |
| The live comparison is explicitly CONTROLLED fixed5 common PP4 runtime OFF/ON, |
| not a production-exact performance claim. Measurement is split A1/B1 then |
| B2/A2 across two restored HOLDs; dedicated measured whole-block evidence is |
| required before either can contact hosts. Missing budget evidence refuses. |
| See ../../../launch/windows/PP4-COORDINATOR.md for the90min bound, pair-only |
| qualification path and outstanding engine/GB10 gates. |
| |
