GV

GeezerVison

The World Wide Web, restored to 800 pixels of glory.

Home|About|Gaming cabinet|Archives|How it works|Starter sites

On this site

  • Sign in
  • ActionsCI/CD and automation
  • CodespacesCloud development environments
  • All Features
  • Documentation
  • Training
  • Certifications
  • Build with Copilot
  • Customize Copilot
  • Integrations
  • Trending repos
  • GitHub Sponsors
  • Developer articles
  • Enterprises

This view

  • Original page
  • New address

GeezerVison is showing github.com — GitHub - evilbinary/yllm: cpu inference only use 49M mem ,distribute inference

BranchesTags
Open more actions menu

Latest commit

 

History

333 Commits
333 Commits

Folders and files

NameName
Last commit message
Last commit date
docs
docs
 
 
inference
inference
 
 
platform
platform
 
 
serve
serve
 
 
tests
tests
 
 
tools
tools
 
 
.gitignore
.gitignore
 
 
LICENSE
LICENSE
 
 
Makefile
Makefile
 
 
README.md
README.md
 
 
main.c
main.c
 
 
serve.yaml
serve.yaml
 
 

Repository files navigation

yllm

轻量级大语言模型推理工具:最小 49M 内存,支持 GGUF/Safetensors 转换、常驻推理服务与分布式层流水线。

快速开始

make avx2                        # 构建 AVX2 优化版 -> build/avx2/yllm
make chat-avx2 CHAT_PROMPT=你好啊 # 转换模型 + 聊天(未转换时自动从 GGUF 转)
make test-avx2                   # 跑测试

构建标量版:make(不带 OpenMP:make OMPFLAG=)。

支持的模型

转换输入 GGUF / Safetensors;量化 F32/F16/BF16/Q4_K/Q5_K/Q6_K/IQ4XS,以及 PrismML PTQ1_0 三元权重(带 Hadamard 激活折叠)。

架构 模型 说明
llama TinyLlama 1.1B Chat make chat-avx2 默认模型
qwen Qwen2.5-Instruct 1.5B / 7B make chat-qwen2.5-1.5b / chat-qwen2.5-7b
qwen Qwen3 8B make chat-qwen3-8b
qwen35 Qwen3.8-27B Gated Attention + GDN 混合;make chat-qwen3.8-27b-avx2(-cuda)
qwen35 Qwen3.5-0.8B make chat-qwen3.5-0.8b(小模型,内存/CPU 友好)
qwen35 MiniCPM5-2B make chat-minicpm5-2b
qwen35 Ternary-Bonsai-2-27B PTQ1_0 三元 + Prism Hadamard,仅 CPU;yllm convert --gguf 转换
gemma4 Gemma 4 E2B / E4B make chat-gemma4-e2b / chat-gemma4-e4b

多模态(make chat-* IMAGE=<图片> 自动带 mmproj):Qwen3-VL 2B、MiniCPM-V 4.6、Gemma 4V。

推理服务(serve)

OpenAI 兼容 HTTP 默认在 127.0.0.1:8000。角色、部署、ctl、HTTP 等全部命令见 docs/serve-cli.md,架构见 docs/serving-architecture.md。

模型按层切分到多台机器,每台只 mmap/计算自己的层段;机器间每 token 只传激活向量。用法、参数与吞吐模型见 docs/distributed-cpu-inference.md。

目录

inference/   推理内核(platform / log / llf / convert / tokenizer / matvec / engine / dist)
serve/       服务层(frame / node / sock 抽象 + rank / server / router / supervisor / hub)
tools/       命令行工具(dump)
main.c       入口: yllm <convert|check|gen|chat|rank|server|router|supervisor|hub|...>
tests/       单元与回归测试
docs/        设计文档

相关文档

  • docs/serving-architecture.md 常驻推理服务架构
  • docs/serve-cli.md serve 命令速查(转换/部署/ctl/HTTP)
  • docs/distributed-cpu-inference.md 分布式层流水线
  • docs/design-mmap-layer-streaming.md mmap 层流式加载
  • docs/llf-vs-gguf.md / docs/memory-vs-picolm.md 格式与内存对比
  • docs/benchmark.md / docs/qwen35-arch.md

About

cpu inference only use 49M mem ,distribute inference

Topics

distributed-systemsinferencemodelqwen3-8-27braspberry-pi

Resources

Readme
Apache-2.0 license
Activity

Stars

7 stars

Watchers

0 watching

Forks

1 fork
Report repository

Releases

Packages

Contributors

Languages

You are visitor 01402503 — GeezerVison does not store the pages you look up.
Best viewed with Netscape Navigator 4.7 or Internet Explorer 6 at 800×600. No cookies required.