DSpark

DSpark extends DFlash with two heads on top of the same block-parallel draft backbone: a low-rank Markov head that biases each draft position by the token before it, and a confidence head that predicts per-position acceptance probability. Pure block-parallel drafting has no dependency between tokens inside a block, so acceptance decays toward the end of the block -- the Markov head restores that dependency, and the confidence signal indicates how far a block is worth verifying. The draft model subclasses DFlash, so the architecture and training pipeline are otherwise unchanged, and it can be paired with any supported verifier. Serving uses vLLM's own dspark method ("method": "dspark" in --speculative-config).

How It Works

Markov Head

The head adds a low-rank logit bias B = W1 @ W2 to the DFlash logits: W1 embeds the previous block token (verifier vocabulary) into markov_rank dimensions and W2 projects to the draft vocabulary. Three variants are available:

vanilla (default): bias from the previous token alone
gated: bias gated by the backbone hidden state
rnn: recurrent state carried across positions within the block

Setting --markov-rank 0 disables the head, leaving pure DFlash drafting. It must be paired with --no-confidence-head-with-markov, since that option requires a Markov head.

Confidence Head

A linear head predicts each position's acceptance probability from the backbone hidden state, concatenated with the Markov previous-token embedding when --confidence-head-with-markov is set. It is trained with a BCE term weighted by --confidence-head-alpha.

Sample From Anchor

DSpark defaults to sample_from_anchor: True -- the anchor and all mask positions predict future tokens, producing block_size speculative tokens. See DFlash for details.

Key Parameters

Parameter	Default	Description
`--markov-rank`	256	Low-rank dimension of the Markov logit bias (0 disables the head)
`--markov-head-type`	`vanilla`	Sequential head variant: `vanilla`, `gated`, or `rnn`
`--enable-confidence-head`	enabled	Attach the per-position acceptance head
`--confidence-head-with-markov`	enabled	Feed the Markov previous-token embedding into the confidence head
`--confidence-head-alpha`	1.0	Weight of the confidence-head BCE term

All DFlash parameters (--block-size, --max-anchors, --num-layers, ...) apply unchanged.

Pretrained Models

Pretrained DSpark speculator models are available on HuggingFace from the RedHatAI speculator models collection:

Verifier	Speculator
`zai-org/GLM-5.2-FP8`	`RedHatAI/GLM-5.2-speculator.dspark-preview`

To train your own, see examples/train/dspark_qwen3_0_6b_sharegpt_online.sh.

Research & Citation

DSpark is based on research from DeepSeek: arXiv Paper

@article{cheng2026dspark,
  title={DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation},
  author={Cheng, Xin and Yu, Xingkai and Shao, Chenze and Li, Jiashi and Xiong, Yunfan and Qian, Yi and Zhu, Jiaqi and Ma, Shirong and Zhang, Xiaokang and Ye, Jiasheng and others},
  journal={arXiv preprint arXiv:2607.05147},
  year={2026}
}