Do I have to patch omp to use a local model?

No source changes are needed: omp judgment already targets a replaceable HTTP contract, so pointing TYPESAFE_BASE_URL at the local :8787 bridge plus a placeholder key reroutes decisions.

How will I know if the local bridge has died?

It silently falls back through tiny, smol, default, and session models, getting slower without an error; check systemctl status or /healthz to confirm the bridge is alive.

What GPU and memory budget does this assume?

The tested setup uses an 8 GB Blackwell card with CUDA builds, about 1.5 GB resident for two checkpoints, and CPU fallback at hundreds of milliseconds; preload avoids 7-10s cold loads breaching the 10s timeout.

Is local Laya accurate zero-shot on my decisions?

No. The base checkpoint scores near chance on typed decisions out of the box; the repo treats Laya as a fast fine-tunable base and points to the RLCD notebook for real accuracy.

What does install.sh actually set up?

It creates a venv with uv and Python 3.12, installs CUDA torch, laya, and mcp, downloads the 2.3 GB weights to ~/laya-omp/hf-cache, and writes and enables a laya-systemone systemd user service. Adding --configure-omp also writes the omp config and sets providers.judgmentProvider to typesafe.

Which model names can I point the bridge at?

TYPESAFE_DEFAULT_MODEL can be laya, laya-english, laya-multilingual, or laya-typed-decisions. The default preloads english and multilingual; set LAYA_PRELOAD_TYPED_DECISIONS=1 to keep all three checkpoints resident.

What happens when TypeSafe requests fail?

omp silently falls back through tiny, smol, default, and the current session model, so a dead bridge gets slower and dumber rather than erroring. Check systemctl --user status laya-systemone or the /healthz endpoint to confirm it is alive.

CommunityJev Agents & MCP

Laya Deploy

Reproducible deployment turning local open-weight Laya into the judgment backend for oh-my-pi, so small checks run on your own GPU.

See highlights

Open github.com

Stars: 0
Language: Python
Published by: Community
Author: Pidbid
Last verified: Oct 2026

Pidbid/laya-deploy

User repository on GitHub

View on GitHub

Deploy the Laya System 1 decision engine as a local judgment backend for omp (oh-my-pi): TypeSafe-compatible /v1/systemone bridge, MCP tools, systemd unit, installer, verification probes and the pitfalls hit on the way.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

把开源 System 1 决策引擎 Laya 变成 omp (oh-my-pi) 的本地判定后端，并让 omp 里所有「小判断」都跑在本机 GPU 上（20–400 ms），不再走云端 TypeSafe。. 不改 omp 源码：omp 的判定后端本来就是可替换的 HTTP 接口。. 实测环境：omp/18.2.6、laya 0.3.4、torch 2.11.0+cu128、transformers 5.17.0、mcp 2.2.0、RTX 5050 8 GB（Blackwell / sm120）、Ubuntu / Linux 7.0。. omp 把「agent 对自身状态做的小判断」（eval 里的 judge()、auto thinking 档位分类、Smart unexpected-stop 检测、git AI staging）统一收敛到一个 judgment 抽象，默认后端就是 TypeSafe 的 System One，而 base URL / key / 模型名都可覆盖：.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Thin adapter serving /v1/systemone with strict choice, noul, and score type-echo semantics omp validates.
Install script provisions venv, CUDA torch, weights, and systemd service with CPU and offline options.
Four verification probes cover contract, omp judge(), MCP tools, and shell-independent .env loading.
Optional MCP layer exposes decide, routing, guard, moderation, and triage tools to agents.

Watch out

Base checkpoints are near-random zero-shot on typed decisions and multilingual temperature is unfitted; plan calibration and the RLCD notebook before production use. 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.

More Jev resources like this

agent-guard

Claude Code PreToolUse hook that screens Bash, file edits, and tool calls with Jev risk, policy, and allow-ask-deny judgments before execution.