Qwen3.8-27B MTP draft vocabulary — Cyrillic variant (59,539 ids)
A draft-head token list for HyperQwen's MTP speculative decoding on Qwen3.8-27B. It is the shipped 40,960-id list plus every whole-UTF-8 Cyrillic row of the vocabulary (18,580 rows), built with HyperQwen's own tool. Speculative decoding is exact: this changes speed, never the output.
The shipped list holds 1 of the 18,580 Cyrillic rows, so on Russian (Ukrainian, Bulgarian, Serbian...) answers the drafter can propose almost nothing and MTP collapses to ~1.4 tokens per step.
Numbers
RTX 3090, vLLM 0.32 (HyperQwen 0.32 port), -fast model, SPEC=mtp CTX=long (fp8 KV, k=3), T=0.7,
8 prompts × ≤700 tokens per language, server-side counters, single pass (±5% run spread):
| draft head | rows | RU tok/step | RU tok/s | EN tok/s |
|---|---|---|---|---|
| shipped list | 40,960 | 1.38 | 55.2 | 107.1 |
| this list | 59,539 | 2.60 | 101.5 | 101.9 |
full head (MTP_DRAFT_VOCAB=0) |
248,320 | 2.69 | 102.8 | 105.1 |
(Maintainer's independent reproduction on their 3090: shipped 1.45 tok/step / 57.5 tok/s RU, full head 2.70 / 100.7 — see gotcha 61 and issue #411.)
Provenance & reproducibility
Built with HyperQwen's own tool, from a public corpus only, so the list is exactly reproducible:
# corpus: aiplatforms/wikipedia-RuDataset (cc-by-sa-3.0), assistant-side Russian rows,
# 16,141 texts / 20 MB / 3.19M tokens
python prepare/build_draft_vocab.py <model-dir> \
--ids prepare/draft_vocab_ids.json \
--add-scripts cyrillic --script-budget 20000 \
--corpus wiki_ru.jsonl --out draft_vocab_ids_cyrillic.json
The whole Cyrillic script (18,579 eligible rows) fits inside the 20,000-row budget, so corpus ranking does not apply — any public Russian corpus yields the identical set. Verified: this rebuild's id set is set-equal to the production list that produced the numbers above.
Held-out coverage (every 10th text, model's own tokenizer — same method as the tool's README):
| corpus | held-out tokens | shipped list | this list |
|---|---|---|---|
| wiki-RU public mix (16,141 texts) | 319,248 | 17.25% (Cyrillic 0.22%) | 99.74% (Cyrillic 100%) |
| real agent RU prompts (8 texts) | 38,544 | 72.55% (Cyrillic 0.14%) | 95.88% (Cyrillic 100%) |
+84% decode on Russian over the shipped list; ties the full head. English within run spread. Background and the Han-cohort equivalent: syv-ai/HyperQwen#196, #295.
Files
draft_vocab_ids_cyrillic.json— 59,539 token ids (the shipped 40,960 first, unchanged and in order, then the Cyrillic rows)mtp_draft_vocab_ids.pt— the same 59,539 ids as a sorted int64 tensor, exactly asprepare/build_draft_vocab.pywrites it next to the model
No model weights here: the head itself is sliced from your own checkpoint in one step.
Use
# in a HyperQwen checkout, on a prepared -fast model directory
python prepare/build_draft_vocab.py models/Qwen3.8-27B-W4A16-AutoRound-fast \
--ids draft_vocab_ids_cyrillic.json
# restart the server; check the boot log / --stats
python prepare/build_draft_vocab.py models/Qwen3.8-27B-W4A16-AutoRound-fast --stats
Rebuild it yourself
python prepare/build_draft_vocab.py models/Qwen3.8-27B-W4A16-AutoRound-fast \
--ids prepare/draft_vocab_ids.json --add-scripts cyrillic --script-budget 20000 \
--corpus any_russian_text.txt --out draft_vocab_ids_cyrillic.json
--script-budget 20000 is above the 18,580 eligible rows, so every Cyrillic row is taken and the corpus
only decides their order — any Russian text gives the same set.
Same thing for Strata
Strata ships an equivalent subset (data/draft_vocab_cyrillic.bin, English/code + whole Cyrillic,
58,963 ids) for Qwen3.8-Flash-Next; it is opt-in via --mtp-draft-vocab. On an RTX 5090 it took Russian
decode from 151 to 212 tok/s (acceptance 42% → 69%).