vLLM and the Speculators library now support three parallel drafting algorithms — P-EAGLE, DFlash, and DSpark — that overcome the sequential bottleneck of traditional speculative decoding. Unlike auto-regressive drafters (e.g., EAGLE-3), these algorithms predict an entire block of candidate tokens in a single forward pass, decoupling drafting latency from speculation length. P-EAGLE uses block sparsification during training; DFlash injects verifier hidden states into the speculator's KV-cache via block diffusion; DSpark adds an autoregressive correction head and a confidence head to reduce wasted verification compute. All three are open-source, production-ready in vLLM, and preserve output distribution exactly via rejection sampling.
Table of contents
1. Introduction2. The Limits of Recursive Drafting3. The Shift to Parallel Drafting4. Under the Hood: Inference & Training Architecture5. Inference Performance6. Production Serving with vLLM and Speculators7. Get Started68 Impressions