GGRun 是 AI Skill Hub 本期精选AI工具之一。综合评分 8.0 分,整体质量较高。我们强烈推荐将其纳入你的 AI 工具库,帮助提升工作效率。
GGRun 是一款基于 Go 开发的开源工具,专注于 gguf、golang、inference-server 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
GGRun 是一款基于 Go 开发的开源工具,专注于 gguf、golang、inference-server 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 方式一:go install(推荐) go install github.com/raketenkater/ggrun@latest # 方式二:从源码编译 git clone https://github.com/raketenkater/ggrun cd ggrun go build -o ggrun . # 方式三:下载预编译二进制 # 访问 Releases 页面下载对应平台二进制文件 # https://github.com/raketenkater/ggrun/releases
# 查看帮助 ggrun --help # 基本运行 ggrun [options] <input> # 详细使用说明请查阅文档 # https://github.com/raketenkater/ggrun
# ggrun 配置说明 # 查看配置选项 ggrun --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量(覆盖配置文件) export GGRUN_CONFIG="/path/to/config.yml"
ggrun (pronounced "g-run", from "gguf run") is my launcher around llama.cpp and ik_llama.cpp. It reads a GGUF's tensor layout against the VRAM, RAM, and per-GPU bandwidth actually on the machine, then places big MoE models across a mismatched multi-GPU setup so the split fits and the server starts without an OOM or a pile of hand-tuned flags. I started it because loading one GGUF is easy, but getting that split right by hand was not.
The project is mainly about three things:
1. running big MoE models that do not fit neatly into VRAM; 2. finding a fast configuration that also stays stable at the context and load I actually want to use; 3. making Claude Code's Ultracode workflows actually usable with a local model: parallel agents, tools, long contexts, research, and no cloud inference for the model calls.
ggrun is not an inference engine. It reads the GGUF and the machine, builds a launch plan for the selected backend, checks that the plan fits, starts the server, and keeps the generated command visible. Unknown flags still pass through to llama-server.
Linux / macOS:
curl -fsSL https://raw.githubusercontent.com/raketenkater/ggrun/main/setup.sh | bash
Windows (PowerShell):
iwr -useb https://raw.githubusercontent.com/raketenkater/ggrun/main/install.ps1 | iex
Then run a local GGUF, download one from Hugging Face, or open the TUI:
```bash ggrun model.gguf ggrun unsloth/Qwen3.6-27B-GGUF --download
ggrun model.gguf --claude-code
This starts the model, points Claude Code model aliases at the local Anthropic-compatible endpoint, and launches the claude CLI when it is installed. The point is to make Claude Code's Ultracode workflows usable with a local model: parallel agents, tools, long contexts, research, and no cloud inference for the model calls.
Context is shared between slots: 1M total context with --parallel 4 is about 256k per request. ggrun lowers the default parallelism when that split would make the individual slots too small, and explicit values always win.
Agent loops resend a mostly-repeated prompt every turn — system prompt, tool schemas, prior turns — so reprocessing all of it each time is the expensive part, not generation. On backends that can shift a transformer context, Claude mode turns on --cache-reuse 256, which reuses prompt chunks that moved after a compaction or context shift, not just an exact prefix match: a compacted 4,506-token prefill went from 45.1 seconds to one processed token in 0.15 seconds in a production test. Hybrid/recurrent and multimodal models that can't shift context that way (native DeepSeek V4, Laguna) get a rolling context checkpoint per slot instead, kept when there's enough host RAM headroom to hold it. Either path, a turn after the first one costs a fraction of what it would cold. Mechanics and the opt-out flags are in docs/usage.md.
Claude Code itself still needs to be installed separately. ggrun replaces its model endpoint and wires the local workflow; it does not make a model with weak tool use behave like a strong coding model. The complete setup and overrides are documented in docs/usage.md.
高性能AI模型推理工具,易于使用
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
经综合评估,GGRun 在AI工具赛道中表现稳健,质量优秀。如果你已有明确的使用需求,可以直接上手体验;如果还在评估阶段,建议对比同类工具后再做决策。
| 原始名称 | ggrun |
| 原始描述 | 开源AI工具:Auto-tuned launcher for GGUF models on llama.cpp / ik_llama.cpp — OpenAI-compati。⭐239 · Go |
| Topics | ggufgolanginference-serverllama-cpp |
| GitHub | https://github.com/raketenkater/ggrun |
| License | MIT |
| 语言 | Go |
收录时间:2026-07-02 · 更新时间:2026-07-03 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。