GGML 库正式推出 v1.0.0 版本,标志着该项目进入新的里程碑阶段 1。此次更新重点在于代码库的清理与现代化,移除了已弃用的公共 API 以及旧式的 DIAG_MASK 操作 1。为确保系统兼容性,RPC 协议版本升级至 v9,以拒绝不兼容的对等节点连接 1。
在功能扩展方面,新版本引入了对可配置 YaRN 截断(truncation)及 FP8 格式的支持 1。同时,开发团队针对多种硬件后端进行了广泛的性能优化与特性增强 1。CUDA 和 Metal 后端实现了少行量化矩阵乘法(few-row quantized matmuls)以及混合专家模型(MoE)的融合计算 1;Vulkan 后端新增了预填充阶段的 FlashAttention 路径,而 SYCL 后端则引入了设备端的 MoE 调度器 1。此外,OpenCL 后端支持分解的 Adreno FlashAttention,WebGPU 后端增加了内核支持,Hexagon 后端的算子覆盖率也得到进一步扩展 1。
The GGML library has released version 1.0.0, marking a significant milestone for the project by cleaning up deprecated APIs and introducing new technical capabilities 1. This update removes obsolete public interfaces and legacy DIAG_MASK operations to streamline the codebase 1. Additionally, the RPC protocol version has been upgraded to v9, ensuring that incompatible peer nodes are rejected during communication 1. The release also adds support for configurable YaRN truncation and the FP8 data format 1.
Significant performance optimizations have been implemented across multiple hardware backends to enhance efficiency in matrix multiplication and attention mechanisms 1. On CUDA and Metal platforms, the update introduces few-row quantized matrix multiplications and fused MoE (Mixture of Experts) operations 1. For Vulkan users, a new prefill FlashAttention path has been added, while the SYCL backend now features an on-device MoE scheduler 1. Further improvements include decomposed Adreno FlashAttention for OpenCL, new kernel support for WebGPU, and expanded operator coverage for the Hexagon backend 1.
评论
还没有评论,欢迎留下第一条。