xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - xPress is a lightweight causal refiner that fixes the independence problem in block-diffusion speculative-decoding drafters, raising draft acceptance length ~30% and end-to-end throughput ~1.3x on Qwen3-8B. It matters because it recovers much of the wasted potential of single-pass parallel drafting without reintroducing sequential token-by-token generation.
- Block-diffusion drafters (e.g. dFlash) emit a whole draft block in one forward pass, but sample each position from its own marginal, so tokens are conditionally independent and often jointly improbable under the target model — causing early rejection during verification.
- xPress reconciles the entire diffusion block at once via parallel refinement, restoring and propagating causal dependencies across draft positions without an autoregressive loop, preserving the speed advantage of one-pass drafting.
- Evaluated on Qwen3-8B across seven math, code, and chat benchmarks: acceptance length up ~30% on average (up to +56%), decoding throughput up ~1.3x on average (up to 1.7x) versus the base dFlash drafter.
- The approach is drafter-side only and lightweight, so gains come without changing the target model or the standard speculative verification procedure.
Sources (1)
xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding
TL;DR - xPress is a lightweight causal refiner that fixes the independence problem in block-diffusion speculative-decoding drafters, raising draft acceptance length ~30% and end-to-end throughput ~1.3x on Qwen3-8B. It matters because it recovers much of the wasted potential of single-pass parallel drafting without reintroducing sequential token-by-token generation.
- Block-diffusion drafters (e.g. dFlash) emit a whole draft block in one forward pass, but sample each position from its own marginal, so tokens are conditionally independent and often jointly improbable under the target model — causing early rejection during verification.
- xPress reconciles the entire diffusion block at once via parallel refinement, restoring and propagating causal dependencies across draft positions without an autoregressive loop, preserving the speed advantage of one-pass drafting.
- Evaluated on Qwen3-8B across seven math, code, and chat benchmarks: acceptance length up ~30% on average (up to +56%), decoding throughput up ~1.3x on average (up to 1.7x) versus the base dFlash drafter.
- The approach is drafter-side only and lightweight, so gains come without changing the target model or the standard speculative verification procedure.