DFlash Deep Dive: Block-Diffusion Drafting, KV Injection, and a Bubble-Free Pipeline

1. Background and origin

DFlash was systematically introduced by Z Lab + SGLang + Modal in a blog post on 2026-06-15, with the goal of pushing speculative decoding on the Qwen3 family up to 6×. It breaks the “serial drafting” ceiling from Ep2 because it reworks both the drafting paradigm and the execution engine.

2. Core 1: Block Diffusion Drafter

Classic drafters are autoregressive and serial — they guess one token at a time. DFlash switches to a diffusion paradigm:

• A single forward pass proposes an entire block of masked future tokens in parallel (default block size = 16, --speculative-dflash-block-size);
• Intuitively, instead of “guessing the next token one by one”, it “fills in a whole block of possibilities at once, then confirms them together”;
• The cost is that diffusion denoising needs multiple iterations, but versus K serial autoregressive forward passes this is still substantially cheaper.

This single move sidesteps Ep2’s bottlenecks (a) serial drafting and (c) in-block sequential dependency — drafting goes from “serial token chaining” to “parallel block emission”.

3. Core 2: Target hidden-state conditioning + KV injection

Parallelism alone is not enough; the acceptance length α is what drives speedup. DFlash’s key trick is to inject the target model’s intermediate hidden states into the drafter’s KV projections (across layers):

• The drafter no longer “guesses blind” — it is conditioned on the target model’s features;
• KV injection aligns the drafter’s proposals tightly with the target context → acceptance rate goes up significantly.

This is where DFlash’s high α comes from, and a major reason it is ~2.5× faster than EAGLE-3.

4. Core 3: Spec V2 engine + overlap scheduler (execution layer)

Most SD work only changes the algorithm; DFlash also reworks the execution engine:

• On a standard pipeline, host-side cleanup / KV allocation and GPU compute leave synchronization bubbles;
• The overlap scheduler runs host tasks concurrently with GPU compute → an additional +33% throughput.

This is what separates DFlash from “pure-algorithm” approaches: it changes both the drafting paradigm and the execution engine.

5. Measured results

6. Summary and next

DFlash’s triple move — block-diffusion parallel drafting + KV injection for higher acceptance + bubble-free pipeline — lifts SD from 2–3× to 6×. Next we look at the alternative route DSpark: instead of diffusion, it uses semi-autoregressive drafting + a Markov head + a confidence scheduler to take a different path.

This is Ep3 of the “Speculative Decoding Notes” series. Prior: The Autoregressive Drafter’s Ceiling; head-to-head: DFlash vs DSpark; next: DSpark Deep Dive.