Phase 3 Re-draft

3Every run, measured on this box

Table 0. Every run, all numbers measured on this 4-GPU box (batch size 1, the same 4-GPU load for every row): rd1 and rd2 per benchmark, plus a base row, the base pipeline (SDAR-1.7B drafter + the trained 128M judge, rd's exact flags without re-draft), where that run exists. Accuracy is items correct, %, with the paired difference from the base pipeline in points (the base pipeline's per-item records are the Phase 1 ones, which this box reproduces token for token). Passes per block and 30B expert loads per token (unique expert loads summed over layers ÷ generated tokens) are counted per item. Tokens per second = total generated tokens ÷ total decode seconds; base tokens per second is on the same items as rd, and speedup = rd ÷ base tokens per second on identical items (item count in brackets). Wall is the run's wall-clock time on 4 GPUs, start to finish, model load included. Reproduction: on this box the base pipeline reproduces the Phase 1 records token for token (HumanEval 164/164, MBPP 257/257, 12/12 on each sampled set), with the same adapter 35805974 and 30B revision c351bbc. Where the base pipeline was decoded on only the first 12 items, the speedup (marked †) pairs rd with the base on those 12 items and is not fully measured: on HumanEval and MBPP a 12-item speedup differed from the full-set answer by up to 13%, and the 12-item base alone has a bootstrap 90% range of −6.5% to +8.0% (HumanEval) and −3.8% to +3.9% (MBPP). Nothing in this table is extrapolated; Phase 1 box numbers are in section 6.

4Results

Table 1. The same measurements arranged against the Phase 2 page. Tokens per second are measured on this box: rd1, rd2 and the base pipeline were all decoded here, and speedup = rd tokens per second ÷ base tokens per second on identical items. HumanEval and MBPP have a full-set base run, so those speedups are fully measured; where the base was decoded on only 12 items the speedup is marked † and is not fully measured (see Table 0 for its uncertainty). Stock, SDAR-1.7B, CoDA-1.7B and Qwen3-0.6B columns are Phase 2 records: accuracy copied from the Phase 2 page, speed measured on the Phase 1 box and listed in section 6, not here. Accuracy deltas are against stock. Highlighted: best in the row.

Benchmark	Measure	Stock 30BPhase 2 records	SDAR-1.7B + judgePhase 2 records	CoDA-1.7B + judgePhase 2 records	Qwen3-0.6B + judgePhase 2 records	Base pipelinemeasured on this box	rd1re-draft every pass	rd2re-draft every 2 passes

Tokens per second = total generated tokens (to end of answer) ÷ total decode seconds over the items, timed with perf_counter and CUDA synchronized at both ends, so it includes prompt prefill, every 30B and drafter forward and the KV flush, and excludes scoring and file writing. The fully-measured row pools only benchmarks with a full-set base run and both rd runs finished ({{ALLNOTE}}).

5Figures

Figure 1

Speedup over the base pipeline, measured on this box

Tokens per second ÷ the base pipeline's, both decoded on this box (4 GPUs busy), full pipeline (30B, judge, drafter and re-draft all counted). Higher is better. Dashed line: the base pipeline at 1×. HumanEval and MBPP: rd and base on every item. Values marked † use only the 12-item base run and are not fully measured. Stock and the other Phase 2 drafters were not measured on this box and are not shown. Nothing in the figures is extrapolated.

Figure 2

Accuracy by benchmark

Items correct, %, one panel per benchmark with its own y-range; Stock, CoDA and Qwen3 are Phase 2 records, the base pipeline is its Phase 1 records (reproduced token for token on this box), rd1 and rd2 are measured here. Whiskers on rd1 and rd2: 95% CI of the paired difference from the base pipeline, 2,000 bootstrap resamples, centred on the point.

6Extrapolated to the Phase 1 box (8× RTX PRO 6000 Server)

EXTRAPOLATED — based on the extrapolation using the measured numbers on 4× GPUs for this dataset for this model, using reference from the runs from this model (the base pipeline measured on the Phase 1 8× box).

Formula. Phase 1 box tokens per second for rd = the base pipeline's tokens per second measured on the Phase 1 box (Phase 2 records) × the speedup measured here on this dataset (rd ÷ base tokens per second, identical items, Table 0); the Calculation column shows the two factors. † = the base pipeline was decoded here on only the first 12 items of this benchmark, so the speedup is estimated by the box-ratio method (rd's tokens per second over all its items here × the base pipeline's Phase 1 box tokens per second ÷ its tokens per second here, both on those 12 items): not fully measured, and a 12-item estimate can be off by up to about 13%. Δ = extrapolated rd − the base pipeline measured on the Phase 1 box. × stock = extrapolated tokens per second ÷ the stock 30B's, measured on the Phase 1 box (Phase 2 page). The 8-GPU wall time replays this box's per-item times over 8 shards through the Phase 2 per-component box ratios and cost model, so it is modelled, not measured. The base pipeline row is not extrapolated: its Phase 1 box value was measured on that box.

Benchmark	Setup	Calculation base tok/s (Phase 1 box) × speedup (here)	Tokens / s on the Phase 1 box	Δ vs base pipeline (tokens / s)	× stock (Phase 1 box)	Wall on 8 GPUs

Reference: Phase 2 records, measured on the Phase 1 box

Table 2b. Tokens per second from the Phase 2 page (× = speedup over stock). These are measurements on the Phase 1 box, not extrapolations; SDAR-1.7B + judge is the base pipeline.

Benchmark	Stock 30B	SDAR-1.7B + judge (base pipeline)	CoDA-1.7B + judge	Qwen3-0.6B + judge

7Progress

The {{NTOTAL}} runs in the order the script runs them (it alternates which setup goes first on each benchmark); base is the base pipeline decoded in full on this box, which makes that benchmark's speedup fully measured. Finish times are estimates for this box, which has 4 GPUs; “running” means the benchmark is still in progress.

Setup	Benchmark	Items done / total	Finish (IST)	State

8Where the speed comes from

Everything here is measured on this box, per architecture: passes per block and 30B expert loads per token (unique expert loads summed over layers, per generated token) do not depend on the machine; drafter forwards per token counts every SDAR-1.7B forward (first draft and re-drafts); tokens per second is wall time on this box. In brackets: stock 30B from the Phase 2 records. Speedup is rd ÷ base pipeline on identical items; † = only the 12-item base run.

Benchmark	Setup	Items	Passes / block (stock)	30B expert loads / token (stock)	Drafter forwards / token	Tok/s this box	Speedup vs base

Baseline. The base pipeline was decoded here under the same 4-GPU load as rd: every item of the benchmarks that have a full base run (HumanEval and MBPP so far, plus any marked base in Table 0), the first 12 items of the others. On this box it reproduces the Phase 1 records token for token (HumanEval 164/164, MBPP 257/257, 12/12 on each sampled set), with the same adapter 35805974 and 30B revision c351bbc, so the work is identical across boxes. The 12-item baselines are small: per-item speed varies a lot, and a 12-item pairing on HumanEval or MBPP would have been off by up to 13% against the full answer.

Cost model. The Phase 2 cost model enters only the modelled 8-GPU wall times in section 6; no measured number on this page depends on it.

Timing: per-item seconds are measured from before prefill to after the last forward with CUDA synchronized at both ends; no polling or sleeps are involved.