User repository on GitHub
Jev-style structured decisions on DiffusionGemma: the example server from vLLM PR 57250
- Forks
- 12
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Djev
Example server that exposes Jev-style typed decisions and grounded span reads by seeding DiffusionGemma canvases and reading one denoise step per slot.
- Stars
- 113
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- mmastrac
- Last verified
- Oct 2026
About this project
A discrete diffusion model denoises a whole canvas per forward pass. If the canvas is seeded with the answer's fixed text and only the answer slots are left as noise, one denoise step gives a distribution over each slot. These extraargs fields (vllmxargs on the OpenAI server) expose that:
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Seeds the diffusion canvas with fixed answer text and leaves only answer slots as noise for one-step slot distributions.
- Serves /v1/systemone with noul, choice, score, span, and spans shapes plus an OpenAI-style chat route with pass-through.
- Supports depends_on staging, ask_if gating, alone reads, windowing for long texts, and an optional thought channel.
- Reports standard error over noise draws with confidence and coverage per span, retrying low-confidence reads up to three seeds.
Watch out
Example from a vLLM branch, not stock vLLM; constrained reads and engine-side sampling need open engine changes and behave differently otherwise. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Yes. It runs in front of a vLLM built from the structured-reads branch with canvas seeding, pinning, step caps, and exact logprobs. Async scheduling is automatic, and synchronous execution supports full-width canvases only.
On stock vLLM, engine-samples fails with a 400 because seed is rejected, while constrained mode is ignored. With the engine changes, constrained reads run the unembedding over labels only, and engine samples fan out N noise draws in one request.
Span reads seed the reply canvas with a label plus blank, pin the rest, and decode with the text's own token ids, so the answer is a substring by construction. Confidence is the weakest normalized probability along the span plus its end.
Declare depends_on to read a question in a later stage, ask_if to ask only when another answer is among listed values, and alone for an isolated read. Span questions run beside label reads and cannot join depends_on or ask_if.
Setting think to N in the schema lets the model write up to N tokens in its thought channel before the read, and the answer slots then condition on that thought. The noise draws of one decision share a single thought, and diagnostics.thought returns the text, its token length, whether the model closed the channel itself, and the generation time.
A text with more distinct tokens than one read can score, 128, is split at sentences into windows with one sentence of overlap. The prompt always carries the whole text so the model never answers from a fragment; windows are read in parallel, the best confidence and coverage wins, and a disagreement is settled by one further one-step choice over the candidates.
Yes. It treats a first system or developer message holding a JSON object with a questions key as a structured read, and passes anything else through to vLLM unchanged, streaming included. GET /v1/models also passes through, so a router can front this port, but the model field must then be the served name rather than an arbitrary label.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython45GitHub stars
Drop-in service that exposes Upstage Solar models through the System One shape, accepting state plus noul, choice, or score questions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
hunkim/solar-mini4-jevPython
hunkim/solar-mini4-jevPython38GitHub stars
Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
FeiLiuEM/open-medical-jevPython
FeiLiuEM/open-medical-jevPython62GitHub stars
Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.
Jev Demos & Experiments#community#jevbest#catalog-import
40GitHub stars
Open research toolkit that encodes an image once and answers multiple typed choice, noul, and score questions from a shared distribution, plus a local 0.6B text-decision model.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
IamBusy/OpenJev-VisionPython
IamBusy/OpenJev-VisionPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.