Qwen3.8-27B-ASCII-Condensed-DFlash2-GGUF

Qwen3.8-27B ASCII-Condensed DFlash2 GGUF

A vocabulary-condensed Q2_K DFlash2 draft model for the ASCII-condensed Qwen3.8-27B targets. It is not a standalone language model: pair it with a matching condensed target to speed up token generation in llama.cpp using DFlash instead of MTP.

Why

Speculative decoding makes Qwen3.8-27B generate tokens faster: a small draft model proposes several tokens at once, and the target verifies them in one pass. This repository provides the DFlash2 draft model for the ASCII-condensed Qwen3.8-27B targets from bsaleh03's work. Pair it with one of those targets and you can run DFlash instead of llama.cpp's built-in MTP (multi-token prediction), which is faster: in controlled text benchmarks DFlash beat MTP by 18-25% at long context. The draft is condensed to the same ASCII-focused vocabulary (129,006 tokens, down from 248,320) so it matches the target's tokenizer and fits in limited VRAM.

Quick start

Transform a compatible Qwen3.8-27B ASCII-condensed target using bsaleh03/Qwen3.8-27B-ASCII-Condensed, then pair it with this draft and start a server:

llama-server \
  -m /path/to/compatible-Qwen3.8-27B-ASCII-Condensed-target.gguf \
  -md ./Qwen3.8-27B-ASCII-Condensed-DFlash2-Q2_K.gguf \
  --spec-type draft-dflash \
  --spec-draft-n-min 0 --spec-draft-n-max 4 --spec-draft-p-min 0.0 \
  -c 90200 -np 1 -b 4096 -ub 512 -fa on \
  -ctk q4_0 -ctv q4_0 -ctkd q4_0 -ctvd q4_0 \
  -dev CUDA0 -ngl all -devd CUDA0 -ngld all -fit off \
  --jinja --reasoning on

Tested on an RTX 5060 Ti 16 GB with upstream llama.cpp commit 5b59b83f (b11051).

File

Compatible targets

This draft is built to pair with the ASCII-condensed Qwen3.8-27B targets converted using the methods described in bsaleh03/Qwen3.8-27B-ASCII-Condensed. It will pair with with the Qwen3.8 target when all of these hold:

• The target uses the same 129,006-token ASCII-condensed vocabulary.
• Retained tokens share the same old-to-new ID mapping and row order.
• Tokenizer merges and special-token IDs match this draft.
• The target keeps the Qwen3.8-27B architecture this DFlash2 model expects.

Vocabulary policy

The vocabulary was reduced from 248,320 to 129,006 tokens (BPE merges from 247,587 to 128,449). The conversion keeps:

• Normal tokens whose decoded bytes are ASCII.
• Invalid UTF-8 fragments needed for byte fallback.
• Every non-normal token, including all Qwen special tokens.
• All 256 byte-fallback tokens.

Non-ASCII text still works through byte fallback, but usually costs more tokens and may perform worse than the full-vocabulary model. This artifact is aimed at English, source code, JSON, shell commands, and logs.

The DFlash mask token was remapped from source ID 248070 to condensed ID 128756.

Only two model tensors were changed: selector_predecessor.weight and selector_successor.weight. These are DFlash2's vocabulary-shaped selector codebooks, with one row per token. The conversion gathered the 129,006 retained rows using the same old-to-new token mapping as the target, without dequantizing or requantizing them. The other 79 DFlash tensors were copied byte-for-byte; only the tokenizer metadata, merge table, and special-token IDs were rewritten to match the condensed vocabulary.

Benchmarks

Controlled text-only SPEED-Bench on an RTX 5060 Ti 16 GB, 90K context, Q4_0 KV:

These measure speed and acceptance, not general model quality. Full setup and raw data are in BENCHMARKS.md.

Validation

• Draft tokenizer metadata is identical to the matching condensed target tokenizer.
• All tokenizer *_token_id fields were remapped through the same old-to-new ID map.
• All 79 non-vocabulary draft tensors are byte-identical to the source Q2_K GGUF (Anbeeld/Qwen3.8-27B-DFlash2-GGUF).
• Every retained row of selector_predecessor.weight and selector_successor.weight matches the source.
• All 276 special tokens and all 256 byte-fallback tokens are preserved.
• Greedy target-only and DFlash runs produced byte-identical output for the tested code prompt.

Sources and attribution

• Qwen/Qwen3.8-27B
• incoai/Qwen3.8-27B-DFlash2
• z-lab/Qwen3.8-27B-DFlash2
• Anbeeld/Qwen3.8-27B-DFlash2-GGUF
• bsaleh03/Qwen3.8-27B-ASCII-Condensed
• DFlash reference implementation

License

Apache License 2.0. See LICENSE.