Learn · Speed Hacks · Frontier Speculative & Distributed Optimization
DFlash & DFlash-2: Diffusion Speculative Decoding
In plain English
Standard speculative decoding generates draft words one by one like a fast typewriter. DFlash uses diffusion to paint an entire 5-token candidate sentence block in parallel, speeding up generation by up to 3x! 🎨⚡
What is actually going on
DFlash and DFlash-2 (Diffusion Speculative Flash Decoding) replace sequential autoregressive draft models with a lightweight diffusion process. A block of multi-token drafts is generated concurrently in continuous latent space and verified in a single target model forward pass, overcoming the draft autoregression bottleneck.
Why it matters
Traditional speculative decoding is limited by the latency of the draft model's sequential autoregression. DFlash generates multi-token candidate trees in parallel, raising acceptance rates and overall tokens/sec.
Another way to think about it
Stamping an entire pre-printed paragraph on paper all at once instead of typing letter by letter. 📄
Check yourself
How does DFlash generate candidate speculative tokens?
- Sequentially, one by one
- In parallel via a block diffusion process
- By querying Google Search
- By skipping attention layers
Show answer
In parallel via a block diffusion process — DFlash generates multi-token candidate blocks simultaneously using diffusion rather than autoregression.
Speculative decoding alters the mathematical probability distribution of the target model.
- True
- False
Show answer
False — Rejection sampling guarantees the exact same output token probability distribution as standard greedy or temperature decoding.