1Status: what ran, what did not

Method

Runs

Table 0. Every run, in the order it was decoded. Items done is the number of per-item records on disk. Wall is the run's wall-clock time on this 4-GPU box, start to finish, model load included. Finished is the time of the last record (IST).

Setup	Benchmark	Items done / total	Wall (4 GPUs)	Finished (IST)	State

2MEASURED on this 4-GPU box

MEASURED — nothing in this section is extrapolated. Batch size 1, one decode process per GPU, the other GPUs of this box busy decoding. Speedup = rd tokens per second ÷ base tokens per second on the same items, both decoded here; it is the raw ratio of two measurements made hours apart on a shared host, so it contains the host-speed drift of section 3. Accuracy is items correct, %; Δ is the paired difference rd − base on the same items with a 95% bootstrap interval. Passes per block and 30B expert loads per token (unique expert loads summed over layers ÷ generated tokens) do not depend on the machine; drafter forwards per token counts every SDAR-1.7B forward (first draft and re-drafts).

Table 1. Base pipeline, rd1 and rd2 on every finished benchmark.

Benchmark	Setup	Items	Accuracy %	Δ accuracy vs base, pts [95% CI]	Passes / block	30B expert loads / token	Drafter forwards / token	Tokens / s	ms / token	Wall (4 GPUs)	Speedup vs base (raw, measured)

Table 2. The base pipeline on this box against the Phase 1 records, for every benchmark it ran on. Identical = the generated token ids of an item equal the Phase 1 box's record for that item. The last column is this box's time ÷ the Phase 1 box's time for the same items (both measured), which is how much slower this 4-GPU shared box runs the same work.

Benchmark	Items run here	Token-identical to Phase 1	Base tokens / s here	This box ÷ Phase 1 box (time)

Figure 1

Speedup over the base pipeline, measured on this box

Raw tokens per second ÷ the base pipeline's, both decoded on this box, full pipeline (30B, judge, drafter and re-draft all counted). Higher is better. Dashed line: the base pipeline at 1×. The host-speed drift of section 3 is not removed. Nothing in this figure is extrapolated.

Figure 2

Accuracy by benchmark

Items correct, %, one panel per benchmark with its own y-range. Stock, CoDA-1.7B and Qwen3-0.6B are Phase 2 records; the base pipeline and rd1, rd2 are measured here (the base pipeline reproduces the Phase 1 records token for token). Whiskers on rd1 and rd2: 95% bootstrap interval of the paired difference from the base pipeline, centred on the point.

3Host-speed drift on this box

Why the measured speedup is not a clean number.

Table 3. Host index H per run, measured from identical prompt-prefill work (rd run ÷ base run, same items). H = 1: the host ran at the same speed during both runs; H < 1: faster during rd; H > 1: slower during rd.

Benchmark	Setup	H, 30B prompt prefill	H, drafter prompt prefill	H, first half of items	H, second half of items

Table 4. Raw measured speedup next to the drift-adjusted one. The adjusted column is an adjustment, NOT a measurement; the project lead does not accept it as a result. It is shown only to size the problem: adjusted = raw × H (30B prompt prefill).

Benchmark	Setup	Raw speedup (measured)	H (30B prefill)	Speedup × H adjusted, NOT a measurement

4EXTRAPOLATED to the Phase 1 box (8× RTX PRO 6000 Server)

{{XSENT}}

Table 5. Every extrapolated cell carries a bold tag. The base pipeline column is not extrapolated: it was measured on the Phase 1 box. The drift-adjusted variant is an adjustment of an extrapolation and is not accepted as a result.

Benchmark	Setup	Base pipeline tokens / s, Phase 1 box (MEASURED there; stock in brackets)	rd tokens / s, Phase 1 box (raw)	× stock (raw)	rd tokens / s, drift-adjusted (× stock)	Wall on 8 GPUs

5Formulas and how it is measured

6How to fetch and run it

Fetch

Run the numbers yourself

File map

Paths are relative to the repository root and are the same in the Hugging Face dataset and on the GitHub branch.

Path	What it holds