sawrus/gemma

最近记录的来源活动: 2026年4月28日 10:46

SkillsMP 收录数据更新: 2026年11月9日 14:22

已收集 skills: 11

GitHub 星标: 0
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/sawrus/gemma.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Gemma Local Agent Runtime Makefile-driven tooling for running Gemma 4 locally on Apple Silicon as an OpenAI-compatible server for Codex and opencode. ## Quick Start Edit `.env` and set `HF_TOKEN` when Hugging Face requires authentication. ```sh make install make model-download make model-repair make serve ``` `make serve` requires a completed local download and checks for `.models/gemma4-26b-a4b-it-tq-2bit/manifest.json`. It will fail fast instead of letting the backend download the model on the first request. Lazy remote loading is available only through `python scripts/serve_openai.py --allow-remote-model`. By default it exposes a short public model alias through an OpenAI-compatible proxy and keeps the raw `mlx_lm.server` backend on `GEMMA_BACKEND_PORT`. `make model-repair` fixes known MLX metadata mismatches for the selected profile, including the Gemma 4 26B TurboQuant 2-bit config bit-width. Default server URL: ```text http://127.0.0.1:8080/v1 ``` Use any non-empty local API key placeholder, for example `local`. ## opencode Custom Provider Start the local server first: ```sh make serve ``` Then add a credential in opencode with `/connect`, choose `Other`, use provider id `gemma-local`, and enter any non-empty key such as `local`. Add or merge this provider block into your opencode config: ```jsonc { "$schema": "https://opencode.ai/config.json", "model": "gemma-local/gemma4-26b-a4b-it-tq-2bit", "small_model": "gemma-local/gemma4-26b-a4b-it-tq-2bit", "provider": { "gemma-local": { "npm": "@ai-sdk/openai-compatible", "name": "Gemma 4 26B TurboQuant (local)", "options": { "baseURL": "http://127.0.0.1:8080/v1" }, "models": { "gemma4-26b-a4b-it-tq-2bit": { "name": "Gemma 4 26B A4B TurboQuant 2-bit local" } } } } } ``` The public model name is `GEMMA_MODEL_ALIAS` from `.env`. The alias proxy maps that name to the local absolute path before forwarding requests to `mlx-vlm`, so opencode does not need to know the filesystem path. Sending the Hugging Face id directly to a raw MLX backend would trigger another download. This repository already has an `.opencode/opencode.json` for agent workflow settings. Merge the `provider`, `model`, and `small_model` keys into that file or into your global `~/.config/opencode/opencode.json`. ## Commands ```sh make help make benchmark make model-repair make lint make test make test-e2e ``` Real model checks are explicit because they download and load the model: ```sh make test-real make benchmark-real ``` See [docs/gemma-local-agent/README.md](docs/gemma-local-agent/README.md) for configuration details.

正在加载仓库 README…

这个仓库中的 skills

软件开发工程师网络与计算机系统管理员数据库架构师

3 个职业分类 · 已分类 100%