IT之家 11 月 14 日消息,开发者 Georgi Gerganov(X 账号为 @ggerganov)昨日(11 月 13 日)在 X 平台发布推文,宣布开源 AI 推理项目 llama.cpp 与底层机器学习引擎 ggml 的库版本迈入 1.0.0。
IT之家查询公开资料,Gerganov 是开源项目 llama.cpp 的创建者,也是 ggml、whisper.cpp 等本地 AI 基础软件的关键开发者。他在个人 X 简介中将自己标注为 Hugging Face 工程师,GitHub 资料所在地为保加利亚索非亚。
llama.cpp 是一个主要用 C/C++ 编写的开源 AI 推理项目,核心作用是让大语言模型在个人电脑、服务器等不同硬件上运行。它既可以用于本地部署,也可以用于云端服务,目标是降低部署复杂度,并充分利用设备的计算能力。
本次更新最大的亮点就是明确接口兼容性规则:
| 项目 | 库 | 定位 |
|---|---|---|
| llama.cpp | libllama | C/C++ 大模型推理库 |
| ggml | libggml | 底层机器学习张量库 |
两者今后都会借助主版本号升级,明确标记 API 或 ABI 不兼容的变化,让依赖这些库的应用可以识别升级边界。对于开发者而言,可以从版本号上,更容易判断某次升级是否可能要求自己的应用同步修改。官方给出的版本规则是:
| 版本位 | 什么情况下更新 |
|---|---|
| MAJOR,例如 1.x.x → 2.0.0 | ./include 中公共 C API 浮现破坏性变更;公告同时明确,API / ABI 不兼容会借助主版本升级体现 |
| MINOR,例如 1.0.x → 1.1.0 | 向后兼容的能力、后端更新、模型兼容或 API 新增 |
| PATCH,例如 1.0.0 → 1.0.1 | 不改变 API 的 bug 修复 |
在工程质量维度,官方公告称为了支撑持续加快的开发节奏,团队强调了三项投入:
在后端与模型兼容维度,ggml 借助专用后端兼容不同硬件;每个后端都可以编译为动态库,并按需在运行时加载。
llama.cpp 已兼容超过 150 种模型架构变体及主流推测解码方法。团队还将继续加强自动化测试基础设施,并借助英伟达、AMD、高通、Arm、IBM 和英特尔给出的硬件兼容,扩大跨平台测试覆盖。
本文来源:https://www.ithome.com/1/027/019.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
ITHome reported on November 14 that developer Georgi Gerganov (X handle: @ggerganov) posted a tweet on the X platform yesterday (November 13), announcing that the library versions of the open-source AI inference project llama.cpp and the underlying machine learning engine ggml have reached 1.0.0.
According to public information reviewed by ITHome, Gerganov is the creator of the open-source project llama.cpp and a key developer of local AI infrastructure software such as ggml and whisper.cpp. He lists himself as a Hugging Face engineer in his personal X bio, with his GitHub profile location set to Sofia, Bulgaria.
llama.cpp is an open-source AI inference project primarily written in C/C++, whose core function is to enable large language models to run on various hardware such as personal computers and servers. It can be used for both local deployment and cloud services, aiming to reduce deployment complexity and fully leverage the computing power of the devices.
The biggest highlight of this update is the clarification of the interface compatibility rules:
Going forward, both will utilize major version number upgrades to clearly mark API or ABI incompatible changes, allowing applications that rely on these libraries to identify upgrade boundaries. For developers, this makes it easier to determine from the version number whether a specific upgrade might require corresponding modifications to their applications. The official version rule is:
In terms of engineering quality, the official announcement stated that to support the continuously accelerating development pace, the team highlighted three investments:
In terms of backend and model compatibility, ggml supports different hardware through dedicated backends; each backend can be compiled as a dynamic library and loaded at runtime as needed.
llama.cpp is already compatible with over 150 model architecture variants and mainstream speculative decoding methods. The team will also continue to strengthen its automated testing infrastructure and expand cross-platform test coverage by leveraging hardware compatibility provided by NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel.
