llama.cpp b11224：修复 Vulkan 矩阵运算读取缓存切片错误

AI 摘要

llama.cpp 发布 b11224 版本，主要修复 Vulkan 后端在 mul_mat、mul_mat_id 等矩阵运算中读取较大缓存切片（如 KV cache 的跨步视图）时结果错误的问题。修复通过从 nb[2] 推导原地读取张量的批步长、按跨步范围绑定 A/B 缓冲区，并新增 test-backend-ops 的 m_v 参数测试用例覆盖这些场景。该更新提升了 Vulkan 后端在 Intel、NVIDIA 等设备上推理的正确性。

正文节选

<details open> vulkan: fix wrong results when a mul_mat reads a slice of a larger cache (#28956) * vulkan: read the batch stride of an in place src0 from nb[2] A dim01 contiguous tensor can still be a view whose batches are strided by more than ne[1] rows, the first rows of a KV cache for example. Both the mat-vec and the matrix paths read such a tensor in place but passed ne00*ne01 as the batch stride, so every head past the first read the wrong rows. The same applies to src1. The stride now

发布时间：2026-09-28 15:06

抓取时间：2026-09-28 15:36

来源机构：ggml-org