mtp-draft-vocab-ko

Korean draft vocabularies for MTP speculative decoding, one file per tokenizer family, built from Korean Wikipedia and measured before and after on the same machine. FR-Spec-style: the draft head's output projection is computed over a frequency-ranked subset of the vocabulary; the target verifies over the full vocabulary, so outputs are exactly the same.

Status (2026-10-07): first file measured and verified — Qwen3.8. Other families below are not yet measured; nothing is claimed before it is.

Files

file	family	rows	built from	measured
draft_vocab_ko_qwen3.8_en_code_65k.txt	Qwen3.8 (nvidia/Qwen3.8-Flash-Next-NVFP4)	65,536	Korean Wikipedia (668 MiB sample, kowiki-latest-pages-articles, 2026-10-01 dump), CC BY-SA 4.0	2026-10-07: decode 16.9 → 26.1 tok/s (+54%), acceptance 1.33 → 2.16, zero replacement chars, zero Korean-script drift, veneta-bench 12/12 unaffected

Every file: one token id per line, ascending, nothing else — the family's shipped floor kept whole, every byte-fallback id pinned, the rest filled by Korean frequency. No header line: the serving kit's loader parses each line as a bare integer with no comment support, so provenance (corpus, licence, build date) is tracked in the table above instead.

Models supported today

Qwen3.8-Flash-Next (nvidia/Qwen3.8-Flash-Next-NVFP4) is the only model this works with right now — that's the one file above. Everything else is still ahead: other built-in-MTP-head families (GLM 5.3 Flash, DeepSeek V4.1 Flash, EXAONE 4.x) are waiting on their own verification and measurement, and separate-drafter stacks (Llama 3.3 70B + EAGLE-3 and similar) are waiting on a generic vLLM patch we're building now. Please wait rather than assume either works — the full per-family status table is on the GitHub repo.

Use (vLLM, Qwen3.8-Flash-Next, MiaAI-Lab overlay)

MTP_DRAFT_VOCAB=draft_vocab_ko_qwen3.8_en_code_65k.txt

Engine support: vLLM only, today. Ollama, LM Studio and SGLang support is planned, not yet built — llama.cpp (Ollama/LM Studio) recently gained Qwen3.8-Flash-Next MTP support but an equivalent vocabulary-restriction feature is unconfirmed, and SGLang's similar-sounding --speculative-token-map is EAGLE-2-only with a different file format. See the GitHub repo for the full detail.

Measurements

See the full results table and raw run files in the GitHub repository (protocol: same kit's en+code vocabulary as the baseline, same machine, twenty fixed Korean prompts, temperature 0, three repeats, median). Quality gate: zero replacement characters across every generation in both arms, a dedicated Korean-script fidelity check, and veneta-bench's 12 memory-ability cases run against the draft-vocab arm (12/12). A literal byte-identical diff of greedy outputs was tried and dropped as not meaningful on this engine — vLLM's continuous batching changes the floating-point reduction order per batch, so even repeat runs within one arm vary slightly in token count; see results/README.md for the exact example that showed this.

Credit and licence

Method: FR-Spec (Zhao et al., ACL 2025) and MiaAI-Lab's DGX Spark kits for Qwen3.8-Flash-Next (the MTP_DRAFT_VOCAB overlay, the 47k floor, byte-fallback pinning, the language-extension method). We are not first to trim a drafter's vocabulary: vLLM issue #58578 (akapug) and PR #59740 (stecasta, "Context Aware Sparse LM Head") proposed and built it before we did, with a more sophisticated mechanism, measured on the same model and hardware class as ours at different numbers (+15.2%/+22.6% there vs. +54% here) that we don't yet have an explanation for. See the GitHub repo's "Prior art" section for the full comparison and why our work is not a duplicate. Frequency source: Korean Wikipedia, CC BY-SA 4.0, attributed in every file header; the files contain token ids, not text. Everything here is Apache-2.0. No customer or private corpus is used. Source and issues: https://github.com/veneta-ai/mtp-draft-vocab-ko