llama.cpp b12080：CUDA 批量张量读取聚合至映射主机内存

AI 摘要

llama.cpp 发布 b12080 版本，核心变更是 CUDA 后端将批量张量读取聚合到映射的主机内存中（PR #31391），并针对小读取与大批量读取分别采用不同策略（小读取聚合、大读取使用 pinned DMA），同时引入条件分发、RAII 主机注册与 128k 描述符固定缓冲区等优化。该版本同时提供 macOS、Linux、Windows、Android 等多平台多后端（CUDA、Vulkan、ROCm、SYCL、OpenVINO 等）的预编译包。

正文节选

<details open> cuda : gather batched tensor reads into mapped host memory (PR to master) (#31391) * cuda: gather batched tensor reads into mapped host memory Assisted-by: Codex * add comments Assisted-by: Codex * simplify the implementation based on checks that exist earlier in the flow * simplify batched tensor gather indexing * add conditional dispatch * address comments - make host registration raii, better test mindful of ci time, assert instead of return false * cuda : gather smal

发布时间：2026-11-04 06:21

抓取时间：2026-11-04 07:18

来源机构：ggml-org