rwkv-jev
Jev-class decision server on an RWKV-7 1B checkpoint that reads typed answers from a constrained softmax behind /v1/systemone.
Overview
rwkv-jev serves a Jev-style decision model built on an RWKV-7 G1k checkpoint of about 1B parameters from BlinkDL. You POST state and a bounded rubric to /v1/systemone and get a typed answer with a full probability distribution over the exact label set; answers are read from a constrained softmax over the answer tokens, so no prose is generated and output cost is zero. The request and response contract matches TypeSafe's, so JevBench's typesafe adapter works unchanged. The list summary is in Chinese.
The engine runs on the Albatross RWKV-7 backend with true-batch prefill and decode in fp16, one GPU forward per answer position, and a trie mode that enumerates multi-token answer paths. Concurrent requests can be merged into one batch. It needs one NVIDIA GPU with at least 8 GB VRAM and a C++/CUDA toolchain, since kernels are JIT-compiled on first start; a reference CPU backend exists for slow smoke tests. The README includes a JevBench self-measurement on public items.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Under the stated policy, is the action permitted? | — |
| Choice | Choice questions over an exact label set | — |
How it works
- Download the 3 GB RWKV-7 checkpoint
- Start serve.py with the Albatross backend
- POST state and questions to /v1/systemone
- Read the constrained-softmax distribution over labels
Strengths
- Zero generated tokens
- Drop-in Jev-compatible contract
- Batched GPU inference
Caveats
- Needs an NVIDIA GPU and CUDA toolchain
- Small 1B model from temporary training checkpoints
- Weights follow the RWKV licence
Getting started
pip install -r requirements.txt, download the checkpoint, then python serve.py --model models/rwkv-g1k-1b-temp-6003.pth --port 8642 --backend v3a --mode trie --strategy "cuda fp16".At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- HTTP server (PyTorch + CUDA)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
基于 RWKV-7 (g1k-1b) 的 Jev 式零生成决策引擎：受限 softmax 直读概率分布，支持 noul/choice/score…
wh000wh000/awesome-jev-live
Frequently asked questions
What model does rwkv-jev use?
An RWKV-7 G1k checkpoint of about 1B parameters from BlinkDL.
Is it compatible with Jev clients?
Yes, it implements the same /v1/systemone contract.
Can it run on CPU?
Only a slow reference backend for smoke tests; the main path needs a CUDA GPU.
Related projects
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchCExperimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
ProjectresearchPythonReads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.
ProjectresearchPython