llama.cpp 与 ggml 发布 v1.0.0,确立 API 兼容性版本规则
Llama.cpp V1.0.0 Sets API Stability Rules for Local AI Tools
ggml-org 于 11 月 13 日发布 llama.cpp v1.0.0,同时 ggml 张量库也达到 v1.0.0,两个项目今后通过主版本号变更提示源码级 API 与 ABI 不兼容。该版本移除了 35 个废弃的公开 API 函数,将 NextN 投机解码 API 从 llama-ext.h 提升到公开的 llama.h,并新增 FP8 KV cache、MoE 模型主机内存缓存等推理能力。llama-server 修复了 SSRF、CORS 和 DNS 重绑定问题,新增 OpenAI 兼容的 /v1/audio/speech 端点,官方发布标签仅在内部 ggml 副本与已发布 ggml 一致时创建。
推荐理由
AI 评分 70
原文梳理了 llama.cpp 与 ggml 同步发布 v1.0.0 后的版本规则和清理项,读者可据此判断升级是否会破坏现有集成。
主题
正文
ggml-org announced llama.cpp v1.0.0 on November 13, 2026, giving the C/C++ local AI inference project a formal compatibility contract for downstream developers. The underlying ggml tensor library also reached v1.0.0, with the project reporting support for more than 150 model-architecture variants and all mainstream speculative-decoding methods.
Versioning Becomes a Public Contract
Both projects will signal future source-level API incompatibilities and ABI, the binary interface used by compiled applications, through major-version increases. Breaking public C API changes trigger a major bump. Backward-compatible features, backend updates, model support and API additions trigger a minor bump, while bug fixes without an API change trigger a patch bump.
The policy follows 3.5 years of development and growing dependence on the project’s library interfaces. llama.cpp provides multimodal LLM inference through libllama, while ggml supplies the machine-learning tensor layer through libggml. The software uses quantization, optimized backends and portable C/C++ code to target hardware from consumer machines to enterprise servers.
That gives package maintainers and application developers a version signal for assessing whether an upgrade can break their integration. The maintainers’ August 7, 2026 release-process discussion adds a related safeguard: an official llama.cpp release tag is created only when its internal ggml copy is identical to the released ggml copy. The condition is intended to keep release-tag builds compatible with a system-wide ggml installation, the arrangement used by the Debian and Homebrew packages described in the document.
Runtime, Model and Server Changes
The llama.cpp v1.0.0 release notes pair that contract with a substantial cleanup. The release removes its deprecated public API, including 35 deprecated functions, while promoting the NextN speculative-decoding API from the llama-ext.h extension header into the public llama.h header. The move places this speculative-decoding interface inside the public API covered by the new versioning rules.
New inference facilities include OCP 8-bit floating-point (FP8) key-value caches, a host-memory cache for mixture-of-experts (MoE) models with a configurable CPU share, and loading of external PLE or engram data from separate GGUF shards. Model additions cover the Kimi-K3.1-Pro, Ripple and xing4_0 text architectures, drama-3 text-to-speech, and vision support for MageVL-Cowork and Kimi-K3.1.
For llama-server, the release addresses server-side request forgery, cross-origin resource sharing and DNS-rebinding issues, and restricts HF_TOKEN to huggingface.co hosts. MCP tool calls are cancelled when a client disconnects or times out, with notification to the MCP server. A new OpenAI-compatible /v1/audio/speech endpoint extends the server’s audio interface.
The redesigned web interface adds named chat groups and a live-preview Markdown editor. Mobile layouts use device safe areas and full-screen dialogs, while the Android chat editor can open the virtual keyboard.
Hardware Backends and Release Infrastructure
The companion ggml v1.0.0 release removes deprecated APIs and legacy DIAG_MASK operations, adds configurable YaRN truncation and FP8 support, and avoids uploading scheduler inputs that are not read. CUDA and Metal gain few-row quantized matrix multiplication and MoE fusion work. Other backend additions include an Intel Vulkan prefill FlashAttention kernel, a device-side SYCL MoE scheduler, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels and expanded Hexagon operation coverage.
Each ggml backend can be compiled as a dynamic library and optionally loaded at runtime, allowing installations to select hardware-specific code. Pre-release artifacts cover Linux, macOS, Windows, Android and iOS, with Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework providing additional distribution channels whose platform coverage varies.
To support the development pace, the team says Hugging Face Buckets and Jobs are scaling continuous-integration capacity for changes landing on the master branches. Hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel is also enabling more self-hosted runners across major hardware platforms. The maintainers are integrating LLMs into their development cycle as well, while identifying code quality, architectural control and understanding of implementation details as their priorities.
For downstream projects, the release turns compatibility into an explicit development obligation for a fast-moving open-source library. Version rules can communicate breakage; the expanded CI and hardware-runner program helps determine whether routine releases remain dependable across the models and hardware those projects use.