DSpark Deep Dive: Semi-Autoregressive Drafting with a Confidence Scheduler

1. Background and origin

DSpark was open-sourced by DeepSeek + Peking University on 2026-06-27 (MIT license, repo deepseek-ai/DeepSpec). It takes a different route from DFlash: instead of diffusion denoising, it uses semi-autoregressive drafting + a Markov head for ordering and a confidence scheduler to dynamically control the verification length.

2. Core 1: Semi-Autoregressive Drafter + Markov Head

Semi-autoregressive: emits tokens in blocks — parallel across blocks, ordered within a block;
Markov head: injects token-to-token sequential dependency within a block, keeping the in-block generation order consistent (no scrambling).

This contrasts with DFlash’s diffusion paradigm: DSpark sits between “pure autoregressive” and “pure parallel” — more parallel than autoregressive, more controllable than pure diffusion.

3. Core 2: Confidence Scheduler

This is DSpark’s cleverest move:

Instead of a fixed block size, it dynamically decides how many tokens to verify this round based on draft confidence;
High confidence → verify more (swap several tokens at once); low confidence → verify fewer (avoid the waste of feeding a whole block that gets largely rejected).

Versus DFlash’s fixed block size = 16, DSpark’s verification length adapts to “should we trust the draft right now”, pushing closer to α’s theoretical ceiling.

4. Core 3: Parallel target verification, lossless

Like DFlash, DSpark is lossless speculative decoding: the draft block goes to the target model for one parallel forward pass, and rejection sampling guarantees the output distribution is exactly the target model’s.

5. Results and ecosystem

Metric	Value / Status
V4 speedup	57–85% (per DeepSeek / PKU release)
Throughput	+400%
vLLM landing	PR #46995 in progress (reuses sparse-MLA non-causal indexing, `DSparkSpeculator` extends `DFlash`, Triton non-causal SWA kernel)
Multi-backend	Supports SGLang / OpenInfer simultaneously

6. Relationship to DFlash (preview)

DSpark changes “draft ordering + verification-length scheduling”; DFlash changes “draft parallelism paradigm + execution engine” — they are on different layers and orthogonal, hence composable. That is exactly why OpenInfer can run both paths on Qwen3-4B. Full comparison: DFlash vs DSpark.

This is Ep4 of the “Speculative Decoding Notes” series. Prior: DFlash Deep Dive; head-to-head: DFlash vs DSpark.