Structured-decision trick speeds up DiffusionGemma inference 3-10x with one forward per request

bodonoghue85 · x · 2026-09-17

Developer mmastrac sped up Google's DiffusionGemma 3-10x by borrowing Jev's ideas; the implementation is merged into his diffgemma repo (PR #21).

How it works: diffusion models already operate on probabilities—by pinning the structured tokens of the response in place, most denoising work can be skipped. A JSON question schema in the system message turns a /v1/chat/completions request into a single scored denoise forward: the answer template is seeded into the canvas with label slots as noise, and label distributions are read from logits at each slot. Every question costs one forward and each slot conditions bidirectionally on the whole template, modeled after TypeSafe's Jev API (choice/score/yes-no with probabilities and confidence). The author suggests the approach could be ported into vLLM.

Original post →

More from Infra

Infra channel →