VibeVoice-ASR-DFlash2-Drafter-AWQ-W4A16-ASYM

A DFlash 2 block drafter for VibeVoice-ASR (7B) (Ar4ikov/VibeVoice-ASR-AWQ-W4A16-ASYM), for speculative decoding in vibevoice.c. It proposes 8 tokens in one pass; vibevoice.c checks them in one pass of the model and keeps the ones the model agrees with, plus one of its own.

The check is exact: every checked row is computed with the arithmetic of the model's own one-token decode step, so the transcript with this drafter is byte-for-byte the transcript without it — words, timestamps, speakers. The drafter only changes how many passes it takes.

One thing changes with --draft: vibevoice.c then attends with flashinfer (--attn auto picks it), because flashinfer checks several rows in one pass over the cache where the default fa2 walks it once per row. The transcript is then byte-for-byte that of --attn flashinfer without a drafter — on this model the default's words, with a timestamp 10 ms apart on a 32-minute file. --attn fa2 --draft keeps the default's transcript bit for bit, and gains less on long recordings.

This is the drafter with its projections stored as INT4 (AWQ, compressed-tensors pack-quantized, W4A16 asymmetric, groups of 128): 0.55 GB instead of the BF16 checkpoint's (Ar4ikov/VibeVoice-ASR-DFlash2-Drafter). vibevoice.c loads the codes as they are. Accepted drafts on the held-out traces: 5.268 tokens per block, against 5.272 in BF16.

Use

Needs vibevoice.c with DFlash 2 support: branch dflash2 (PR #48), in the next release. The model and this drafter in one download: Ar4ikov/VibeVoice-ASR-AWQ-W4A16-ASYM-DFlash2 (the INT4 drafter in drafter/, used without --draft).

vv_cli --model ./VibeVoice-ASR-AWQ-W4A16-ASYM --audio talk.wav --draft ./VibeVoice-ASR-DFlash2-Drafter-AWQ-W4A16-ASYM
vv_cli serve --model ./VibeVoice-ASR-AWQ-W4A16-ASYM --draft ./VibeVoice-ASR-DFlash2-Drafter-AWQ-W4A16-ASYM --slots 4

It drafts for any VibeVoice-ASR (7B) checkpoint of the same model — microsoft/VibeVoice-ASR, scerz/VibeVoice-ASR-4bit — since it reads only the model's hidden states, embedding and LM head; it was trained on the AWQ checkpoint's transcripts, so that is where it drafts best. On NF4 (bitsandbytes) and INT8 weights only --draft-check fast pays: their decode step has no multi-row twin, so an exact check would cost a step per row, and vibevoice.c declines the drafter there with a pointer to fast.

Results

vibevoice.c b72be15 (branch dflash2), RTX 3090, VibeVoice-ASR-AWQ-W4A16-ASYM, greedy decoding, decode tokens per second:

Plain = the same binary without --draft, both with flashinfer attention, which --draft picks by default; --draft-check exact with the check width it measures. Against the default without a drafter (fa2): 3.65x on the 2-minute file, 2.78x on the 32-minute file. The same clips and files as the BF16 drafter's card.

The drafter

• DFlash 2: 5 Qwen3-style layers (hidden 3584, 28/4 heads, intermediate 9472), 831M parameters, block 8.
• KV injection: for every position the model has processed, the outputs of its layers 1, 7, 13, 19 and 25 are concatenated, projected (fc) and normalized; every drafter layer turns them into keys and values. The drafter reads the model's own view of the audio and of the transcript so far.
• Two-tap dynamic convolution around attention and MLP (per-row kernels predicted from the row), and a candidate selector that re-ranks the model head's top 16 per row with a pairwise predecessor/successor term (rank 256).
• The embedding and the LM head are the model's, frozen. The drafter scores a draft vocabulary of 32768 ids (draft_vocab): the ids its training transcripts use, plus every stop id.
• Accepted drafts on 43 held-out clips' traces: 5.268 tokens per block of 8 (the selector's, with every earlier draft right).

Training

Self-distillation: the drafter learns what this model writes, timestamps and speaker ids included. tools/dflash in vibevoice.c holds the pipeline:

1. ~390 hours of audio, 8228 training clips and 43 held out, from LibriSpeech, FLEURS (8 languages), SOVA, AMI, earnings calls and VoxConverse — audio only; the datasets' own text is never used.
2. The model's own greedy transcripts of every clip (vv_dflash_data gen).
3. Traces: every position's token, its role and the five tapped layers' outputs, replayed through the runtime (vv_dflash_data trace).
4. train.py (PyTorch, flex attention): anchors at generated positions, rows 1..7 scored against the next 7 tokens with weights e^(-k/4); cross-entropy of the head plus the selector's; AdamW, cosine schedule, BF16.
5. awq_drafter.py: activation-aware scales where they fold into the drafter without changing it (the MLP input, up → down, v → o per KV channel), then a clip search on every projection; INT4 codes stored as compressed-tensors.

Limits

• vibevoice.c only: the checkpoint keeps the published DFlash 2 names, but the KV injection, the selector and the draft vocabulary are as vibevoice.c implements them (docs/DFLASH.md).
• CUDA only (the CPU path and Metal decode without the drafter). Greedy only.
• Audio the corpus covered little (Mandarin, Russian) keeps fewer drafts; the runtime then falls back to plain steps and stays near plain speed.

License

MIT, like VibeVoice.