| # STOP: measured primitives do not qualify an exposed production win |
|
|
| All values below are labelled. No live request, Spark access, pair work or |
| production-K2 full-model replay occurred. [Frozen PP2 inputs](pp2-frozen.json) |
| were read with git show from e1befc9, never refitted. [Cost model](estimate.py) |
| and [machine-readable result](cost-estimate.json) preserve unknown costs as null. |
|
|
| MEASURED on x86host's solo bc62df2a at225W, required miniconda CPython3.13, |
| Torch2.12.1/cu130 and Triton3.7.1; graph-replayed primitive timings: |
|
|
| | Batch | Depth4 top32 median/p95 ms | Depth5 top32 median/p95 ms | Depth5 torch argmax median ms | |
| |---|---:|---:|---:| |
| | B1 | .143155/.143258 | .155443/.155546 | .130970 | |
| | B2 | .141926/.142029 | .177357/.177459 | .133325 | |
| | B3 | .150221/.150323 | .187699/.187786 | .133318 | |
| | B4 | .166912/.167110 | .208589/.208691 | .134144 | |
|
|
| These time already-conditioned synthetic logits, not Markov evaluation, the |
| current production fused selector, K2 backbone, real TP2 exchange or eager |
| sampled serving. B1 incremental selector cost versus this torch microbenchmark |
| is .024474ms; B4 is .074445ms. Neither grants replacement credit against the |
| production selector. The producer uses two launches/position and misses the |
| absolute .05ms/round goal at every measured shape. Its prototype model hook |
| still runs the original greedy selector and materializes extra Markov logits; |
| there is no whole-K2 or exposed-pair timing for that path. |
|
|
| MEASURED B1depth5, full target vocabulary129280, actual prepared standard |
| rejection kernels on the same solo: full q-aware verifier p95 .014029ms versus |
| point-mass verifier .013005ms; sparse1280-byte D2H .003859ms, H2D .004307ms; |
| dense-q zero+scatter .004506ms. These exclude target model/processor time, |
| CPU validation, slot index-copy and GB10 differences. Five toy-law tests pass |
| at family alpha1e-5; the numerical test is not a hardware timing transfer proof. |
|
|
| ESTIMATED framing: B1depth5 is1416B/reply including app header, sparse payload, |
| semantic header/mask and CRC; B4 is5412B. At the previously measured40Gb/s link |
| negotiation, B1 serialization is .0002832ms, or .0005664ms charging two serial |
| copies at40Gb/s. This is shape arithmetic, not a network-latency measurement. |
|
|
| Adding the full producer, full q-aware verifier, two transfer proxies, dense |
| expansion and two wire copies gives **.182813ms** in known B1 proxies. This |
| charges full verifier cost rather than only its delta, gives no greedy credit |
| and does not assign a value to missing work. Summing p95 operations is a |
| conditional cost model, not a measured p95 of a pipeline or a statistical upper |
| bound. Unknowns include extra Markov work, eager launches, full vocabulary peer |
| staging, CPU CRC/q/proxy work, target slot copies/TP synchronization, GB10 work, |
| quantizer differences from PP2 and future sampled-policy occupancy. |
|
|
| PP2's pooled nominal top32/.7 fixed5 expectation rises from3.091813 to3.305514 |
| retained tokens/round; paired delta interval .175370–.254434. With its archived |
| 48.08ms fixed5 policy price (not an observed matched mean), the pooled lower |
| interval allows2.727129ms total added cost. After known proxies,2.544316ms |
| remains for unknowns. Family lower-interval remaining budgets are2.320812 and |
| 2.369660ms; the16–32K proxy is tighter at1.365677ms. PP2 used float64 normalized |
| q, whereas this implementation samples emitted dyadic24 q; its acceptance is |
| not newly replay-qualified. There is no150K acceptance capture. |
|
|
| For the requested B1 short/32K/150K view, the following are **ESTIMATED |
| sensitivities**, not forecasts. Each transplants the relevant fixed5 replay's |
| relative token-count gain onto a different historical adaptive W8 time/acceptance |
| population. The32K row uses the16–32K proxy;150K additionally extrapolates pooled |
| anchors no longer than47K. Changing the unknown-cost assumption changes the |
| answer substantially. No actual expected live rate is reported. |
|
|
| | B1 context | Historical adaptive tok/s | Conditional tok/s with unknown extra0ms | With extra1ms | With extra3ms | |
| |---|---:|---:|---:|---:| |
| | Short |53.56|58.01 (+8.32%)|56.52 (+5.52%)|53.74 (+.34%)| |
| |32K proxy|62.64|65.31 (+4.25%)|63.71 (+1.71%)|60.74 (−3.03%)| |
| |150K extrapolation|55.23|58.78 (+6.42%)|57.34 (+3.82%)|54.66 (−1.03%)| |
|
|
| The0ms column illustrates omitted costs; it is not the decision model. Actual |
| net rate and150K acceptance remain null. The current prototype fails the |
| single-launch/absolute timing objectives and lacks real K2/pair/GB10/lifecycle |
| qualification. Preserve production; finish those gates before the separately |
| pre-registered [69-minute coordinator window](LIVE-AB-PLAN.md). |
|
|
