[KongchangAI]
Modelblock drafter / DFlash2块草稿器

Block Drafter

A draft model used for speculative decoding that generates multiple candidate tokens in parallel in units of blocks for verification by the main model. Common implementations include DFlash2 and others.

Timeline (last 90 days)

Oct 25

On code generation tasks, the block drafter accepts on average about 7 tokens per step, while on open-domain prose the acceptance rate drops to about 2.5.

Unverified50%
Sep 22

Block drafters emit tokens at multiple positions in parallel with a single forward pass, at the cost of slightly lower draft quality compared to autoregressive drafters.

Unverified50%

All Facts (2)

Source Articles